diff --git a/Hyper ParameterTuning/n11/README.md b/Hyper ParameterTuning/n11/README.md new file mode 100644 index 0000000000000000000000000000000000000000..b04fba5acedbe02e2407f325254f0f177831cbc6 --- /dev/null +++ b/Hyper ParameterTuning/n11/README.md @@ -0,0 +1,13 @@ +# n11 Hyper Parameter Tuning + +This folder stores RF-guided hyperparameter tuning outputs for n11. + +Structure: +- `/runs/`: copied training/generation run directories referenced by phase1/phase2 score tables. +- `/_rf_tuning/`: tuning manifests, score tables, reports, and final recommendations. + +Models currently included: +- tvae +- tabsyn +- tabdiff +- ctgan diff --git a/Hyper ParameterTuning/n11/ctgan/_rf_tuning/final_recommendation.json b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/final_recommendation.json new file mode 100644 index 0000000000000000000000000000000000000000..ef38c72f5a18b0fd921eeaa67ec2c738637d66e2 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/final_recommendation.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7db55acd4b6a22334b421ffea9eecbd16cd64742a51134d0bad178d1c1318ae +size 3313 diff --git a/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase1_manifest.json b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase1_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8387b447366cb255e9835275af11167e0e72548f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase1_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed2ba5ad2b2f1dde8669d5507e9c7154e8b99027a71b2b521bf11c96474ddf8e +size 3819 diff --git a/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase1_scores.csv b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase1_scores.csv new file mode 100644 index 0000000000000000000000000000000000000000..8201526421c1644599e8f43305def7e9bb6c650a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase1_scores.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c151689e09718b076584b5beadfb7b0325c08a43aafed32aa69c3c47ccd23e3 +size 3367 diff --git a/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase1_summary.json b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase1_summary.json new file mode 100644 index 0000000000000000000000000000000000000000..e7e6c4e6ea2edba4a60fd477b07488fabe7b433b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase1_summary.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fcd98079e4d7e8a448bc6bbba29c5a9b7ec1ef88b301b9229314cc7abe735ed +size 573 diff --git a/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_manifest.json b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..eda5922e5a2cad2ad5eea87326c6156193e32af1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fdb58e45d41345db04a528b885618fa5977ab48940687b5f3e0441aab573fcf +size 3502 diff --git a/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_report.csv b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_report.csv new file mode 100644 index 0000000000000000000000000000000000000000..708e2b4685b62692c4313154e95eaff988ccb503 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_report.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:755fd2f311511fc4d549214017b83bfebd279538e5551d33b3236370ce5a24c5 +size 928 diff --git a/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_report.json b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_report.json new file mode 100644 index 0000000000000000000000000000000000000000..613221daddb10b5d28903222b60d5991ef2caec6 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91ccc0b7d6877d2946d4ef07d97f4dcd821b06c43e17de340f5715dc00ae0312 +size 3591 diff --git a/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_scores.csv b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_scores.csv new file mode 100644 index 0000000000000000000000000000000000000000..3b976fbd8b76bd54b5047359cb8a0d6abe5f57cb --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/phase2_scores.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:754ebaedbb503281299a7ee8899adeebf360a8e73a4df3ae87c44de0914d623e +size 5602 diff --git a/Hyper ParameterTuning/n11/ctgan/_rf_tuning/sample_manifest.json b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/sample_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d67258cbd36db32669dec2db1b837e439e1b228f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/_rf_tuning/sample_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81edfacc2b5a8f1c1176e45bce7632587f0809a04ada0d79084c049a99b9edd9 +size 633 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..aa22e8365af70239d42626d9f5d5fd11ad213915 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_234320/models_100epochs/ctgan_100epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_234320/ctgan-n11-15215-20260520_235141.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_234320/ctgan-n11-15215-20260520_235141.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..23481218e96d1e9fd9f946887b85c5bb519a6419 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_234320/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=128, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=128, + pac=2, + epochs=100, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_234320/models_100epochs/ctgan_100epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_234320/models_100epochs/ctgan_100epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..f98019fa2e7b6b9caeeaf335a9011010e7703e2d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15f6cef7f01ad1c5245381dd4b5ace7f5100333afb48daadaa1a4862c8eafa64 +size 243 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/ctgan-n11-15215-20260520_235141.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/ctgan-n11-15215-20260520_235141.csv new file mode 100644 index 0000000000000000000000000000000000000000..8bde73f117c21f8e0be132b805614c72edc1c94f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/ctgan-n11-15215-20260520_235141.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9985b74bcf70eebc91f403889c202cf427ab18b3509815583761001d4235effa +size 2881394 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/gen_20260520_235141.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/gen_20260520_235141.log new file mode 100644 index 0000000000000000000000000000000000000000..217aad8743a50a017f66fa766a4a31c1e1a62c73 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/gen_20260520_235141.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:213a0bd6efde79490b19c0c6da68713713c85ed880b65ce14927910e8546a8c5 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/models_100epochs/ctgan_100epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/models_100epochs/ctgan_100epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..3d046d607b7320dc2b0f1dbab99af6da325c6dd0 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/models_100epochs/ctgan_100epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc00424d5e606f41028638bc3260105afb333a69acc2d9d0c2af9e868dd045c9 +size 1309923 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/models_100epochs/train_20260520_234320.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/models_100epochs/train_20260520_234320.log new file mode 100644 index 0000000000000000000000000000000000000000..e2cef103470cb1148e3093cc7a5e3aa5dec653f6 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/models_100epochs/train_20260520_234320.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f21fc489fefde084eb3ed023991360880b14891d95ab490593e339f726d13d9e +size 20950 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9e733846740fdbe387c91f1f8e7310ed9875b582 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8dbe319ee6aef48dbe2f4bc930c17bf57da221dab98c7b9e3978094e8e092186 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..89449ec3a0b0c7786f0abbfca7d6beb5ef584c9a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:005d0ac41db5a140ba37ffc97d9825fcf8dac13a1016ebe78b786cb2b565ad5b +size 2473 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..f4d4135060e490c818d27668d3d8fa3f0696ad37 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d124f76b4f6dac706871b06d5127d709d6120e92b1905d81b24b71fe71ce6f48 +size 911 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..60f4961c3aced8c191d4b53fa252e35bcab565fe --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a66fa26b69ff8eaf39fdfbfcf711abc7f8c204295329b1d57c09ef6749e456d9 +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..cf821b230a834b597f32e12e66e93899bad2985f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:186bc4d0b96b7b899f6a8c54534e722e653e9c50273e475ccd5c7e5db7ba15d9 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_234320/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..cd40914b03c6083d6d2d37c6f675da73f40cf82e --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_235258/models_100epochs/ctgan_100epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_235258/ctgan-n11-15215-20260521_000102.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_235258/ctgan-n11-15215-20260521_000102.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ca24d79e1e4bcc9d2aada95e332a1c06008a033b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_235258/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=128, + generator_dim=(32, 32), + discriminator_dim=(32, 32), + batch_size=128, + pac=2, + epochs=100, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_235258/models_100epochs/ctgan_100epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260520_235258/models_100epochs/ctgan_100epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..fd8821e6b925afa87b63b132075b6647cb7b44c6 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebe7230a13978f275e8636146d1ca52d58aaa2d0d6198434cd96522359cf8863 +size 220 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/ctgan-n11-15215-20260521_000102.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/ctgan-n11-15215-20260521_000102.csv new file mode 100644 index 0000000000000000000000000000000000000000..4fe1c7e3a78a89191a0a1f424077fb1ce2ea45a3 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/ctgan-n11-15215-20260521_000102.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c92e70886130da7ff4f9f394c4036851fa217f739e775ceb2ecc69f657d22687 +size 2883061 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/gen_20260521_000102.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/gen_20260521_000102.log new file mode 100644 index 0000000000000000000000000000000000000000..d8ce89af715dc33aae809e1d3f20d55784ff849e --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/gen_20260521_000102.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f3fabf93e2691b87336fe1e06583860176bf5a97ddacfe392160dce632cd52b +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/models_100epochs/ctgan_100epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/models_100epochs/ctgan_100epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..71c12ac5ddf6a5fa8d43c75e64eb8826c54eac82 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/models_100epochs/ctgan_100epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61a63c4f5dcf536f610ea5d8c2edc9ac273b9ec83564ff98941ac762e8f2ba4c +size 614499 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/models_100epochs/train_20260520_235258.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/models_100epochs/train_20260520_235258.log new file mode 100644 index 0000000000000000000000000000000000000000..49ecec60a25e0dbd00a42b5be63a2b4fd899ee5c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/models_100epochs/train_20260520_235258.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a2fe99a81ac2a6a1f7cf4680fe322894262ec74de807eb9f74036ef0df05a69 +size 20829 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ff62cc3b6fd0ce413cf695e67b647e7896adb11d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48dc868f6baff2cda27610f37c96116ec8b49a915822788bcd107f9fea9505f5 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..2c10b4c45f231dcc3ed10a5bc3202d3c5295996d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb647b97f7b514d41969a18904ca2d2ec5dc5cc583072d1d863a828bf5d4b80d +size 2461 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..11414eab7c049aa96ba5e0eca85525ff4e5320cb --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7148d34551ebc7f777731024d18938251181f1f52740d5eb78c5efad88640775 +size 911 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..b38a66106b26ddcdbfe1977d8233edde3e579d00 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:719ee6cb8a4f6acbe67075f12ec530bcf309999938b9990c61c844d4b6d9948c +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..96848901588f6118f08f2078bb8b063a0ecd341d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e29d3da476a897362b63eb2fa41ad184118dfe159a71dce28dedfb97b9376aa +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260520_235258/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..935cb33b020d53a311a8616b0dc10642845ecf9c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_000219/models_150epochs/ctgan_150epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_000219/ctgan-n11-15215-20260521_004831.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_000219/ctgan-n11-15215-20260521_004831.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..43cd902e86f6368419c563a3b66b85c7ba1c5bc4 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_000219/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=128, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=32, + pac=1, + epochs=150, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_000219/models_150epochs/ctgan_150epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_000219/models_150epochs/ctgan_150epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..3fd42f71f8767aeb6b791bbf77869078c54a69ae --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df903523d21157c03d8658e1d1ec2dc3efd94811c055df5cc5edc47a2450254e +size 206 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/ctgan-n11-15215-20260521_004831.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/ctgan-n11-15215-20260521_004831.csv new file mode 100644 index 0000000000000000000000000000000000000000..c0fe65ae25aeb8955e565b829dd65cf3b8a72dac --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/ctgan-n11-15215-20260521_004831.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:245aedf3d13ac57b521160d1728a2fd3f7bda4045e85c25e24f057649632b2af +size 2887499 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/gen_20260521_004831.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/gen_20260521_004831.log new file mode 100644 index 0000000000000000000000000000000000000000..c4eb1b77f138326b652047beb34b70cf28b91fdd --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/gen_20260521_004831.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f05e7d2a8dcab3b3ef3e7d2d1d239a04dac6cfbbc9170aa59e9da0b2cff1a5e1 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/models_150epochs/ctgan_150epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/models_150epochs/ctgan_150epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..5d26fde97f0df990ba3d4a7459764863b8581fa7 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/models_150epochs/ctgan_150epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:507d2e9c20e03220677a42d7ebc9fb604a771bdf531b5d3085927206b6396bc9 +size 1311459 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/models_150epochs/train_20260521_000219.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/models_150epochs/train_20260521_000219.log new file mode 100644 index 0000000000000000000000000000000000000000..7c5d3055870352333c0d3f952923fcc7fd78c35a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/models_150epochs/train_20260521_000219.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e7c8ad499957f22e3e651546af477066527697aede3c5948e146d8830704be4 +size 30414 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..90c21d6dadc71960664300177e010ec10c589012 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8edbd243f0740a89dbd87e5dce3b10a20da59ecbf061063b5071f9642af48ba +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..c57f7c225d9b4d756d97786f7467569a2cf9efda --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2879232048456211a9084ecc942606f4cdd1707b03bd0fb8cc07b3c9d51f4494 +size 2456 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..48eeb6dfd0b00d92dab265eac956cd5d6fc70a86 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3b6f26b9d55a71323aa522aed30d9096d315f62e7e35bf1683930a5099be2b0 +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..0c5752700bcf3006fbf5aa7128b7bad6ea6c3039 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e843e49a1dd9b77351095bc33a1476455ad20f9aaccf026cfeb69494957792c +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..dacea7cd7dff4158172ec2e1b1c3ed919a7098b0 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3a3a71df4df2e52ad3e676a48be56ef592e70b01be5e1e1ae4eb18217093c57 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_000219/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..9089b510b8a8e9c3b126d701889efc38a68fb331 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_004950/models_75epochs/ctgan_75epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_004950/ctgan-n11-15215-20260521_005549.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_004950/ctgan-n11-15215-20260521_005549.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..952a2287e82b2d1ccdaee6e57931c3aa828ea2e9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_004950/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=128, + pac=8, + epochs=75, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_004950/models_75epochs/ctgan_75epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_004950/models_75epochs/ctgan_75epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..34ca1674b10ea8680f15d3ffa6a043a55844bbcc --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0325cd127bf2f686d5040a526142d12d2efdee7ffe9633a94bb2f5383ccb9635 +size 241 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/ctgan-n11-15215-20260521_005549.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/ctgan-n11-15215-20260521_005549.csv new file mode 100644 index 0000000000000000000000000000000000000000..b7ecc5b1b8d171a727474304982591141efd7eb1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/ctgan-n11-15215-20260521_005549.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b82a0c6a20e0274d3cdfa72b2c7a0127964de9fb809ad09d42c481cf7a5def53 +size 2881953 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/gen_20260521_005549.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/gen_20260521_005549.log new file mode 100644 index 0000000000000000000000000000000000000000..77d60e83a233540589095aabb030078b5495ec90 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/gen_20260521_005549.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb7aa3639d9db679cbd859cdb8089332729caea7d7f63b1e9583ac23d9cca2cb +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/models_75epochs/ctgan_75epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/models_75epochs/ctgan_75epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..0de82b31d45d0cad32af487c4211b8bda7c00ab6 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/models_75epochs/ctgan_75epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54829b71d2ebdaddbf41c2576fb0d0ebf7a669f6f1cfb0faa985de7a5846d63a +size 1030861 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/models_75epochs/train_20260521_004950.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/models_75epochs/train_20260521_004950.log new file mode 100644 index 0000000000000000000000000000000000000000..45836006975090418a0bf72f29a8e4f2cda77202 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/models_75epochs/train_20260521_004950.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:701a359645c0db156727e8c9ae96cd12a0d93eebfbb4acf1de6a4220dd29d74d +size 16218 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..3287c705c2251be29bc4b639b413aa6c2dcb0955 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a0e252ba10f4c1ed1e6d6bc88af1b11648f8a1be845656f05a604a1ec16af36 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..5a32c3742745adbd0ad97b4daef1a765ac15ae07 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdb7bb2ce2946df887b6cfdd2ce27e877d64fff0dd5ad0682e56b50d21208e14 +size 2468 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..6e1550efe51fde06b28731c289243884ace3e8e5 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3df7cb601868c311433c7fc6813d26a74f03120e74e5a9c8a786ac9026ecf52 +size 908 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..c5983cf2a80e01a71fe4f18a9f94ad37b892fe54 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e74db0ac501e5c2f282119ae83cb7fa78a3779f0b847be62187a07d5e05fdf49 +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..32d7c82cad6b10a6a3a7a3f3e5fddc672de0ea52 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:646d59838ad9075b2eaaf9f52f37baacc9e1eeae63d5b24ffbfcce0079a49dab +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_004950/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..98724011f9e244742868bc6ac7877e39a932bb82 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_005705/models_350epochs/ctgan_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_005705/ctgan-n11-15215-20260521_012359.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_005705/ctgan-n11-15215-20260521_012359.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..91903a2b7b9fd896b20c803bf7ab4d51a0290bdc --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_005705/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=32, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=128, + pac=1, + epochs=350, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_005705/models_350epochs/ctgan_350epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_005705/models_350epochs/ctgan_350epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..9d2f1c16939780e80ff6d5863a7e38d563e88ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fcee68a532f811c0a4358e5326cc885ea2f1817ab6c35d2d3d0f1a5f5afab97 +size 224 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/ctgan-n11-15215-20260521_012359.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/ctgan-n11-15215-20260521_012359.csv new file mode 100644 index 0000000000000000000000000000000000000000..58fbd1ad4d036332fca8428e9cda4985e15094ec --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/ctgan-n11-15215-20260521_012359.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1417379cfb6214993147f2b0f3de8ce9106b72e4463c44d4a678947405198b4 +size 2890374 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/gen_20260521_012359.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/gen_20260521_012359.log new file mode 100644 index 0000000000000000000000000000000000000000..477f8093e84def74ac6fc89f9e1c6244b9f4ef73 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/gen_20260521_012359.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e19f74cfd2474495b1cc4391001a8546033fda02f9b6db089332d967b4e9e488 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/models_350epochs/ctgan_350epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/models_350epochs/ctgan_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..e11fa20bfdb0379a1643c4ba75e7dc9b3929976b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/models_350epochs/ctgan_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a66e3ca7f9c5cb2713a132b0e4d4b65161abd88ef445f9a120bf63001422f50 +size 1079267 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/models_350epochs/train_20260521_005706.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/models_350epochs/train_20260521_005706.log new file mode 100644 index 0000000000000000000000000000000000000000..0c5031ac756287179411d15d461ab8e7b50dc440 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/models_350epochs/train_20260521_005706.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddfdf61829a06787a3fb52ac7481b695f1e644116ced667bbe912992b34a56b6 +size 67906 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0ef1dcd3facd9e11988c216d8ecd7c4069bf71fa --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8ac2050bb8d5769090887187d9216beb1513b16406d8a61f7d002beb4378cbe +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..90b6106878edc455eb8f2440a00da88738b544fb --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c10618620d3520575fdf9654c6956a2cc3a78acdff5b9706a7efe81affad5ff +size 2464 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..bbfa847069bf373d7050bb268d37a2b1301c51e5 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:142ed5d75ee62968aa844fef1ec94e5b68bd4b430f5d4555700bfbf83829c83e +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..02f0a254db970aab9ca65015fe7f98fdc7ea1bc3 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46f0eaf6bf99280a58c5c64ec370620beb6f30a7e483b62bed334bd70efce297 +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2043a9db04b7538988e6346cb0bd904f104d62cb --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e05e4fd24cb191f3d3628462304421e464f47cbfd6481b2155406a7ecc76dc3 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_005705/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..7ea85d80f375e75dfc60719a3e6821075d66972b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_012519/models_50epochs/ctgan_50epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_012519/ctgan-n11-15215-20260521_014040.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_012519/ctgan-n11-15215-20260521_014040.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..1526b1f308de9204d8ed93fd05094a4ad1559462 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_012519/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=32, + generator_dim=(128, 128), + discriminator_dim=(128, 128), + batch_size=32, + pac=2, + epochs=50, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_012519/models_50epochs/ctgan_50epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_012519/models_50epochs/ctgan_50epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..32f0296a7bbfaedb52f2717ee1328b5987d48b82 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fb02cd37c6d1454bf24f92817943d0384084f94c245bb66b5b6152bd5440423 +size 204 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/ctgan-n11-15215-20260521_014040.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/ctgan-n11-15215-20260521_014040.csv new file mode 100644 index 0000000000000000000000000000000000000000..4dd01800fdb44c16e85c8ec2c263854d43e6b2e3 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/ctgan-n11-15215-20260521_014040.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:092c1079e4bb8390b8ba59f787a2155f8862486c27c060de7c81138e1615402d +size 2887935 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/gen_20260521_014040.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/gen_20260521_014040.log new file mode 100644 index 0000000000000000000000000000000000000000..da66e8dfc807b449161cc4d2a078a0a54e5daea4 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/gen_20260521_014040.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e8185098e7c6461cee8d87d1f66ffe8465a7f64cb9782e72feca342967811fc +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/models_50epochs/ctgan_50epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/models_50epochs/ctgan_50epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..4a5602490ed5ace3f7efb0d829d1ccedb77efeda --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/models_50epochs/ctgan_50epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c147e94151eadc37c8bd9ccd48cfa66590e5b0436cbb2c55b0c2f8677f5d99b +size 721677 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/models_50epochs/train_20260521_012519.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/models_50epochs/train_20260521_012519.log new file mode 100644 index 0000000000000000000000000000000000000000..8807e113a3b5249bba7e95bb018332dcd1aa4419 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/models_50epochs/train_20260521_012519.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f4152a52a960ab8aef066eff8288f7903d4f7080b763f5da42cd1154c0d1dfb +size 11585 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6b0d79b482ef05e188f30f0ba059a69fb29bae30 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30762063baa311af60870575244f43635ee7d0f9125b07cf2d6229ab0538e10f +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..0f7fd3ba25db52f4a66c42630575587f847484f7 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b6e84ab7003e0be50bd8f55011536a428ec798d67b9625522c72e03e493d867 +size 2451 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..5db44f77cec252998f55e5b1b0785203c3128883 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1e41b124289226a7c712f2fadee3b4a5c81442e8ceda13ec571604c77670a83 +size 907 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..838d84702358885fe64cb16a2e03b550286e9133 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f68c2a37b7c8d8b68ae12cc2c5a469a733becc85fe28d547404b1f86e5e8d18b +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2dbe1efabe31b0d399a9ab55c12293bd65d331d2 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa4331e3078423dffb35765ca449d3c4d936a008427a4559f7dda3ea96cba8d9 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_012519/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..21d54af874b7277bbb8a78abfe252bac1055d1ae --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_014159/models_300epochs/ctgan_300epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_014159/ctgan-n11-15215-20260521_031242.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_014159/ctgan-n11-15215-20260521_031242.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ffae2dded1be8628af8da48ecfc7fff866e54bc1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_014159/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=32, + pac=8, + epochs=300, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_014159/models_300epochs/ctgan_300epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_014159/models_300epochs/ctgan_300epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..a16abde9157415957e6b0024eb8584a72b58102a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b71d96c10bb54b0b3bae3ab21e073867ad2352d35cfd73633166f16198d18d1 +size 241 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/ctgan-n11-15215-20260521_031242.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/ctgan-n11-15215-20260521_031242.csv new file mode 100644 index 0000000000000000000000000000000000000000..b8fd57832fa50f88339e89dbb82e67ae7a425a9d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/ctgan-n11-15215-20260521_031242.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5032039aefd4ac0a4cbe4a016d88cc30efd3d46791755b3d1540d9bfcb0134a +size 2884741 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/gen_20260521_031242.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/gen_20260521_031242.log new file mode 100644 index 0000000000000000000000000000000000000000..fa09ff9b0a9d4969b408e623171d0b2ce0bc3d38 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/gen_20260521_031242.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de12aadee0d15e60df9d2d86b82fad6c43f140801457fb3d344dfca8ffe4c7c4 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/models_300epochs/ctgan_300epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/models_300epochs/ctgan_300epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..8cc612f01a93deb0b1e4dc8ffa6c55aaf5cb3716 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/models_300epochs/ctgan_300epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb17cad19709cac957934be28c16e1d49c48eef2bee277c5b3f335b73d35c2a4 +size 1037859 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/models_300epochs/train_20260521_014200.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/models_300epochs/train_20260521_014200.log new file mode 100644 index 0000000000000000000000000000000000000000..cf48634e85c3d57e5b7ed9fddcdf6ee715d26c12 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/models_300epochs/train_20260521_014200.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb3833ce7b5cbb8bfefd36b06595a957aa7dcbd9c8b83262e2da7238de66621b +size 59097 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..cac340cd6abb4e8ff239659d5a2764f22bff49c0 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae05f708d6f8d00cf888ae59f3bcc9948ea2cf206d79686c7f179f6cc8795b0c +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..8a45d2a142d69827d1ade6c1c56ab06764b9d284 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e23f675ab05a6fd91f23bf1a1a8ce59f3b8dcc3405ba98e8447674103d2488b +size 2471 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..5b5efff4e2e06a2721d0012e37d3d2c261153ac5 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54d96c5e1f350c856819aa1281a0aec6f64430be9376056ecaf90c01d40ccb2a +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..cf91bcdd8d3a0f5f3ee05e57e9ab5c44b447ba2e --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af536cdf33a54fe02c99b38dc1d9f1f022c76ee2a733189729a161989bda05cf +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..97f72813061796c41b96d4b121d8546e66b321bd --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a35686ef41fcaf83c1fa111baeb553c638ba51109e3740c8f0a072117a9b5178 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_014159/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..5592002a3002b767908dd8a4fac41d06904f50a1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_031401/models_100epochs/ctgan_100epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_031401/ctgan-n11-15215-20260521_032200.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_031401/ctgan-n11-15215-20260521_032200.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ca7738471a29ba9c775ab839c2606cd66c4f6cc9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_031401/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=64, + generator_dim=(64, 64), + discriminator_dim=(64, 64), + batch_size=128, + pac=4, + epochs=100, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_031401/models_100epochs/ctgan_100epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_031401/models_100epochs/ctgan_100epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..e43c061e2658717c1780da084daccebae4c9a194 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:391e8e309c2d45e52a8071438cd28117eb2b19f806817f0f1ebec6cf5b36dfc2 +size 203 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/ctgan-n11-15215-20260521_032200.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/ctgan-n11-15215-20260521_032200.csv new file mode 100644 index 0000000000000000000000000000000000000000..d9f1a39a0e3788551cfc192d4ba5cc359ee1c674 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/ctgan-n11-15215-20260521_032200.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:373fc6fad6e5faaf8565ebb43480eca15bea88cee2ec45c62078b269515ff14d +size 2890856 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/gen_20260521_032200.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/gen_20260521_032200.log new file mode 100644 index 0000000000000000000000000000000000000000..d91f0d09755877f3f7bf04410bd87259e898cd55 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/gen_20260521_032200.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15879fad4e37f6ad170a2bcd289b14235ea538c0e05c2989d3946dc90e5888d7 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/models_100epochs/ctgan_100epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/models_100epochs/ctgan_100epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..1c08756fc2f8bdfe9b7dcbf84e9a0648ce18d89d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/models_100epochs/ctgan_100epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0da9ea0e5cfbdbfba9a231aa6059f652ff533cf5ff1696a5c6dfa9aecf111c9 +size 628643 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/models_100epochs/train_20260521_031401.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/models_100epochs/train_20260521_031401.log new file mode 100644 index 0000000000000000000000000000000000000000..4d9b79dbc144f1dea5ed190ddbd988121b104285 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/models_100epochs/train_20260521_031401.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d97349d121ec940312089ce5c9061e5199c70c8bc7bd24aed36d9de4d089e86d +size 20966 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4cf9ca450a4578b42d52d3c25d79b53ae863d464 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f885b5b478d7d0a9c328f6d73ee97c643852c56565431f1d6aa08b8659bf5f36 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..76dc42734877e60465a52b2fa39189946c30c7f6 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0837cf7e629f7f64a551bcc9689f035b7d3dd39b7b0c5bdee82dd726511814a4 +size 2454 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..3af01e7d09e71f2eabef4fd61f46ee953ceefb6b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:326d07706f026f5896c82063b75a6f7f71ffa56e0fb5e3d20f24e428ad80ffab +size 911 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..4ac08bd0a4c02ecc1b888bdbc0c9f3c59be2b262 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9250031759ffd3c10a5af0617b74c66de620c0a66af9811c3bb0a68c893826b +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d715c472b8a548947fca371315a930ee508c6dd3 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30be57037db6f860f82819478f58a2ade8b20693d33a7babc8ce9370c51d930d +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_031401/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..b40964e39a316ab07c33e210365f377526b3d1c5 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032318/models_50epochs/ctgan_50epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032318/ctgan-n11-15215-20260521_032537.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032318/ctgan-n11-15215-20260521_032537.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..b45733de8fe03fc78db14ecbd9479b214485b058 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032318/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=32, + generator_dim=(32, 32), + discriminator_dim=(32, 32), + batch_size=256, + pac=8, + epochs=50, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032318/models_50epochs/ctgan_50epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032318/models_50epochs/ctgan_50epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..9f564279550d6aa5cdf6d5f3e33b8ce7f6264256 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccaa9a81c5970fb176c969a65f013625e5c436c5204e96e88275e49f9f1f5f08 +size 202 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/ctgan-n11-15215-20260521_032537.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/ctgan-n11-15215-20260521_032537.csv new file mode 100644 index 0000000000000000000000000000000000000000..4d452a98f4126ef9556ef4ffaedaef4f6b523404 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/ctgan-n11-15215-20260521_032537.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f802f49d1cc834bf3d87996ea1fc652571c4854cc1cc08cb2696a6cbee38aa97 +size 2891111 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/gen_20260521_032537.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/gen_20260521_032537.log new file mode 100644 index 0000000000000000000000000000000000000000..4e081ed16624ea2cd3bf6c41e6d27225778e6000 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/gen_20260521_032537.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40c542f501d788242d463c779dd9874126a633a47b901fcbc9b855524ffb4b45 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/models_50epochs/ctgan_50epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/models_50epochs/ctgan_50epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..d8fa4dab430d40783df7128f23b39904b8a8db4e --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/models_50epochs/ctgan_50epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8071d0855437cbfba35eeebd8b1f1c53f1fa6a306d39f5c7f45668312bd8fcc +size 546573 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/models_50epochs/train_20260521_032318.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/models_50epochs/train_20260521_032318.log new file mode 100644 index 0000000000000000000000000000000000000000..e004fd7e1aef120c8e4ff6815fd32ca023f4d080 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/models_50epochs/train_20260521_032318.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:176d98c71723a0c650edb4f7d2dc9a5098b243acef468ea2218b87a1551e1385 +size 11556 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c453bed3dac8d94fa6e5bf747f83c154f1314f20 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc7305f39b9e04c8a68a6e6e6f0f226b11f37233779c4ab33b30aace5910094f +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..e3c235b2459caf2ce0d27170c1c784b08e6c8b1a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dd4a638f658f1ee3aabef45497a36bac4916a613599daa70dc252ad660dfd29 +size 2450 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..ec28e7a65946f552a8c9e09fc141c313a0cdab75 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5af9280d8478aa2903a2fb699390bac335ea01fd846af926f74e129086e80cf1 +size 908 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..2dcbd81eca20d5b3d49544612d7a9beead13dcd5 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d9e946beb592a97e20ad3324cbbc5c511a036516691ed8c4ff209f510e8e74e +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..425f8c345f2244e91765f5c7098e5303476848db --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f7c324f9a4d9f0549287b94fd36213fda4cb612e0db548a3b3eb4005223c123 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032318/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..53075e277c59b5dce1d13d58adff61b51ac2ac56 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032655/models_250epochs/ctgan_250epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032655/ctgan-n11-15215-20260521_040429.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032655/ctgan-n11-15215-20260521_040429.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..7268341f2e307373182d759ba8ae76e592822140 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032655/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=64, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=64, + pac=1, + epochs=250, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032655/models_250epochs/ctgan_250epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_032655/models_250epochs/ctgan_250epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..f32c896ade8f5f2cad58215a451c7ea330a8e5b9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5c4844e4b08cd1f76f98548f74a17241aee46bb8df7d359db8d8a94122fa1ad +size 223 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/ctgan-n11-15215-20260521_040429.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/ctgan-n11-15215-20260521_040429.csv new file mode 100644 index 0000000000000000000000000000000000000000..0b32f65adf7ca9c7fa9105649212e1033904511c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/ctgan-n11-15215-20260521_040429.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f7a8b9e6a48708cad80880daf147b55f3750cb9a8f733a6eca5eac98d107bea +size 2889509 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/gen_20260521_040429.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/gen_20260521_040429.log new file mode 100644 index 0000000000000000000000000000000000000000..ec8e459851049d159cf7e3ca9782c231523fa0d2 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/gen_20260521_040429.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3bfe24cc1e9c7774c0aa6336c01edaeb7502a8ea99fa4a2e09ad00e4d993ceb +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/models_250epochs/ctgan_250epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/models_250epochs/ctgan_250epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..0f437c7bc7f53da04532f0a3b015f321357f952c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/models_250epochs/ctgan_250epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a4f45b4b8b2badefe770f892021929c6c24f51021c827c45317255b9ff2276d +size 1155619 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/models_250epochs/train_20260521_032655.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/models_250epochs/train_20260521_032655.log new file mode 100644 index 0000000000000000000000000000000000000000..107a8dff6cbac1dd45a45001380783edb6e61f1e --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/models_250epochs/train_20260521_032655.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1fdaf917f3d9fe814c7c93c7b16248a572fddbad87b5c7c1c72c7a0f3db7c05 +size 49147 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2ad4c0a7455b577d386d77384949321f2a4c8d55 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea2279c45cc77339ebbba008882114d6b1c6251d0b76d1ff77fb96b0b76b70ad +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..7248d78c046c5644b6939d08de267be88778e7dc --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c96966c71be17e31f06af22b2ae924d3fe08fb023bbd3c92b86a18dbd31784d3 +size 2463 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..ecb4a6ca4e060b55bc4c4123b95438dda082c7f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:769bf17a943249c212feab6eb9515c5e0c8a664f622343995649afec0f3bab4f +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..fdb8d9c720688d902ad694b2c02de7c210dd35ca --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f40e98d07701fb150ba7d9e5589b3d1236f5f3657ddaf0b8889b125e1928768f +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..246414bf637ba2c5ab7e3757187ec71471f6aede --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da3590165a6e1458d01da89f8f58b04565555db029f5b14f7918df16ea63cedb +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_032655/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..134e2741412f3f22a60fbeaf017b282d50a33e3b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_040548/models_50epochs/ctgan_50epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_040548/ctgan-n11-15215-20260521_041002.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_040548/ctgan-n11-15215-20260521_041002.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..cf4e7621cffb587dc605228467c07600f75ffd56 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_040548/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(32, 32), + discriminator_dim=(32, 32), + batch_size=128, + pac=2, + epochs=50, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_040548/models_50epochs/ctgan_50epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_040548/models_50epochs/ctgan_50epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..c0f5a38ef51afc75226420051165e92f924b57e5 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c887a658e03d034cd2fd552eefa8469d0a557beeaa7cf2db0a9958957fa733b +size 218 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/ctgan-n11-15215-20260521_041002.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/ctgan-n11-15215-20260521_041002.csv new file mode 100644 index 0000000000000000000000000000000000000000..e45ac4f8e5b2976d521e43be2aa764932e41181e --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/ctgan-n11-15215-20260521_041002.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60c143608b55305fadf4d68f7e7ea6dc0dd60b6bd4796381e53da4c8b88c8dc2 +size 2887981 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/gen_20260521_041002.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/gen_20260521_041002.log new file mode 100644 index 0000000000000000000000000000000000000000..894374c56dcf367ab2a0aec0d9e153d902a05977 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/gen_20260521_041002.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d2f57d0d88567297ce9ac60452107405f6ea9d6f99637bfcc6ffed2259d97dc +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/models_50epochs/ctgan_50epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/models_50epochs/ctgan_50epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..23fe5eb0c8815c2562eb1b4d44bf50a53b685a93 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/models_50epochs/ctgan_50epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ade1eef6dc6666614cdeca46b01b2e3cdceb37461cdb4b07dd46044222f5bb4 +size 535501 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/models_50epochs/train_20260521_040548.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/models_50epochs/train_20260521_040548.log new file mode 100644 index 0000000000000000000000000000000000000000..13598e701f4b442ba6d075b3e5680bd8c42d0c15 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/models_50epochs/train_20260521_040548.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5d6db75fbde43b1bb707f942b48f45f30b6ebf4912d787a49013300ccc786d7 +size 11649 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..f329bf3ab3836c1129f03d0b66025a6c7202507b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f61da3e58bc5f570ce2f3edd8e0edb9e99a21aee394c9b2c1c8db92259c16a8 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..ee5028ed7848b71c7929a4f29e1620baee2bfa63 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:269ac3495b8aca969aefa62a0dd64ba1b964d121055c7b81090160bc2a5b4a49 +size 2456 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..650f0fde801220decdd5fd0e736820aa846db4bd --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:367506ee2077846f8f2af02880cceb838a8ddd33789af0b21bb779531cc6470f +size 909 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..b9270dcb0af125468c5b559f0cf07569e12d89b8 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f46ba819611343d67ae05f4919c22b2166e258abd048ca2cb17d6281c8a408c +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..abf56052626017346356976a71d525e23b2b6dcd --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1358c678de885180ee69bd6ad61e804acb94d6a42f7f0f0650fd4ee1fe58106b +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_040548/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..1866fe56ec31a97ff1acb5e19b250b7a70fb9370 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_041119/models_350epochs/ctgan_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_041119/ctgan-n11-15215-20260521_060142.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_041119/ctgan-n11-15215-20260521_060142.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..6205436d3aad433da71932494e1a4e29bb939994 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_041119/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=64, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=32, + pac=2, + epochs=350, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_041119/models_350epochs/ctgan_350epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_041119/models_350epochs/ctgan_350epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..882148f770443dee63683b5de122607001450a5d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:370c32292fb28427d0683d3273e98fcef7a74c71cacccb80f56da531b4c190ed +size 241 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/ctgan-n11-15215-20260521_060142.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/ctgan-n11-15215-20260521_060142.csv new file mode 100644 index 0000000000000000000000000000000000000000..0e449b320c13e0e9c8c6ed314a99953910399983 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/ctgan-n11-15215-20260521_060142.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:621495b6b492aa6a348de2915b73b942c00de2bbcc152280c3799ab84bf65055 +size 2884998 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/gen_20260521_060142.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/gen_20260521_060142.log new file mode 100644 index 0000000000000000000000000000000000000000..1e7b88003a5f9b147dce99378f6b6c994535bff4 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/gen_20260521_060142.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9357067db7ba2fe6dbde75e61385f9dd854d43034453ff66d09116ce5c0660c5 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/models_350epochs/ctgan_350epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/models_350epochs/ctgan_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..2459d2177ca1d89a48571325a66d5f4837c1d45b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/models_350epochs/ctgan_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a48a05bb7d115395ccbfd87b5e6a15ebf12866b2f46bd9db97ec4e925a632ea +size 1158691 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/models_350epochs/train_20260521_041120.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/models_350epochs/train_20260521_041120.log new file mode 100644 index 0000000000000000000000000000000000000000..76dcefd761dbfaa82d261702d85d99065fcafa35 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/models_350epochs/train_20260521_041120.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f2bb2b1f734c0656546e71a14fc3eec4127b590cc2273ff5c40ecefd3fbada3 +size 68955 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..f3c4c9dcfe86cff67625cb5e7759306a2b677225 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77e563ca500c9fbe37675e8e9c00ec9cb653281bac5e169d72ca4a0a89ded947 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..ab1bef9694b5de12ea6f60e2eb0b78ba16c169b9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5b13140c6117c40c6100781a40cbe954b08cffac2e2ccf8f58c31dc2a815943 +size 2471 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..0f6536401946771bcd5efd83068690926052ebf2 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:550b8163c7db964ca5b196722c926b176135c17c644d19db53624cfe5e1f1d3b +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..8f64ad7ef56811c7c290ae1253b948063bee9437 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8aa395434868e485e388269cebdba40d16e8eb5d2748add7d16cb25db701230 +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fd9a55013eaa7804e39ec0ea93145edf74f8cf4b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:770605d6cc459d07002c4866941d79583bde9c0a6439795db226c9e5ea133442 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_041119/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..c9d1c1dd5ef47d5404fd2fbc9406fd87337ffaf8 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_060839/models_350epochs/ctgan_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_060839/ctgan-n11-15215-20260521_063508.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_060839/ctgan-n11-15215-20260521_063508.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..8eebad99e90d36b8fe51ea099ac3b8b774ebf1c5 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_060839/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=128, + pac=1, + epochs=350, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_060839/models_350epochs/ctgan_350epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_060839/models_350epochs/ctgan_350epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..54a11771375bdf8dd8860ea0c85c14d59a142e24 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0648d4d10b40982abe742e275a0c4245bfc8aa5375c0ad44e56e4f124ce25c24 +size 242 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/ctgan-n11-15215-20260521_063508.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/ctgan-n11-15215-20260521_063508.csv new file mode 100644 index 0000000000000000000000000000000000000000..3cdf9d212f6b83bc8ad18d3ea39b7963735cd1eb --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/ctgan-n11-15215-20260521_063508.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:819ac8e3b761c425fb07a055be0899d15d1ec546fd3d8ccecbad5a2d3330c88f +size 2896745 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/gen_20260521_063508.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/gen_20260521_063508.log new file mode 100644 index 0000000000000000000000000000000000000000..d39d9de0660bbe189d82d17e2f2c9a6c961ba361 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/gen_20260521_063508.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58555702a67b6f75fff39bc8d6c01549a5b6d74dbe5c861610de8b3720bb2503 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/models_350epochs/ctgan_350epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/models_350epochs/ctgan_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..d509d84c6261b378fa01203b2e7df04a26fee88d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/models_350epochs/ctgan_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ed2c5b8d90927082a0da15a2d949230acc7c5589b7b6631e773ef102aaa189d +size 1039587 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/models_350epochs/train_20260521_060839.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/models_350epochs/train_20260521_060839.log new file mode 100644 index 0000000000000000000000000000000000000000..399a4c49c9974d12c850ed69d8222701ad73afd3 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/models_350epochs/train_20260521_060839.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72a26eaff99a5b92510d46807243850fe9823b663ab4581ef8dbc18db2b40114 +size 67894 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8f440532817676434f80da3851b23459128dbee1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e83e62888dccfc76f09f48c9a249e5e40fd83f9bd4c9ecef17aabe8d71e73002 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..68fd34d427f7420c8d45e94a8952d988c101456b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c089fc182c410fa62dcb7bdf8650544fe98917affa78dbe95268411f7b59979 +size 2472 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..aa8b42412aa2388c19fa43b5288b346f36c4edd8 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68ea6993d2dd633fc5fb51b20dd6229f25071d9bbf39196e14f4a13f5ba97419 +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..8a8e20b470474c1fd7a0c73133fda6b3db9b09b3 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c7743501f69446bed3379d2c4ceb0c79fd8ce2d43530ecbe94357684976492c +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..1463511d162285f442646757c3441c93071ad28a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec69b8cb3dfe38fad6db1bfef5c2d7f1ca6c43733e436f0f7c95a2040417213b +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_060839/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..ffecd40a40bbba5e071f75dddd3aafcf4ed7aef3 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_063625/models_350epochs/ctgan_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_063625/ctgan-n11-15215-20260521_064946.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_063625/ctgan-n11-15215-20260521_064946.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..fa2ef8d5719402557e3c504b2821c9fbca93278d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_063625/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=256, + pac=1, + epochs=350, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_063625/models_350epochs/ctgan_350epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_063625/models_350epochs/ctgan_350epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..0eacef44482630ebdc253ca22a0092ba6358f994 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:257ff39bdd106f738e8f02b5876772f53b7bf281625bf0aaf47eadfe520aa619 +size 224 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/ctgan-n11-15215-20260521_064946.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/ctgan-n11-15215-20260521_064946.csv new file mode 100644 index 0000000000000000000000000000000000000000..0e4a284033a5e670d496bf93ca86fab16e177277 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/ctgan-n11-15215-20260521_064946.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8076b3004ab83ff80abf0e9878e828dcd04bd7a25c2dd9cb294c0e5a7ad0f230 +size 2888048 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/gen_20260521_064946.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/gen_20260521_064946.log new file mode 100644 index 0000000000000000000000000000000000000000..7d67c57d4f9b4e514fe741ee6f60e6f28dd2eec2 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/gen_20260521_064946.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd91b57d1df8fba1a686b5a8986fd2354465981939975f992139d72644f38495 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/models_350epochs/ctgan_350epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/models_350epochs/ctgan_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..19d13d9b3dfec48e2e53d3101c464d2e13e585fa --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/models_350epochs/ctgan_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09617e33724b2fb6ef7c63f04e11a498e2dcb711fd09f8b4304cc3186bbe2c32 +size 1039523 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/models_350epochs/train_20260521_063625.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/models_350epochs/train_20260521_063625.log new file mode 100644 index 0000000000000000000000000000000000000000..aa656bcef9ba7e64ef6f632f3a064ab8ce9c541a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/models_350epochs/train_20260521_063625.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fb10b8b1d6ea8f7b239d4923569fcef3fd90ceb73995f7248bd8fcc9fd350ad +size 67908 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..794b14786133d9f7ed7cb63595d0a71fff552524 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7972d8b584c620cc5024755cb2dec626843a19b654f5c01670febae3a089f1f +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..38a9780f33599cbca223f7088bd3fef0586bc726 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd2fae3ab140cf50b9ed6c547e10e220c4ee5f2a42e975539e3d82173963cd01 +size 2464 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..a787162e9e225b5b337635b3c52263f13faeb421 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edd8a7d06691a398227687036dfd40b0067fde1d81094a93f56b8fdf9c79fde7 +size 911 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..83b732ed9b852e2bee4978bcb812a3059d804d84 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f0faa2a38e1daf10635c92153c368ddb96ad82ccecdd82a19c1ec8714275a63 +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..042e5e591aa97578d3df9cf463bed07800bc6204 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:521b09323ffdbc6e8f3a35a51b5c4e496d86d63631646bc3bccc8eeed2814e97 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_063625/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..dbd1e64badd134fdaf3d8804f8cbf1e4e906ae80 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_065103/models_350epochs/ctgan_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_065103/ctgan-n11-15215-20260521_074228.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_065103/ctgan-n11-15215-20260521_074228.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..46757369220c8de8dcb8ef03b6346dc1b64b9e89 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_065103/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=64, + pac=2, + epochs=350, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_065103/models_350epochs/ctgan_350epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_065103/models_350epochs/ctgan_350epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..c4ba736e893c2f82084b7b49c975ec4b26030a9f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fca01f77ccce57d565ddf190e0761b6593f8664d5861389ec151afbf71f8b8c8 +size 241 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/ctgan-n11-15215-20260521_074228.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/ctgan-n11-15215-20260521_074228.csv new file mode 100644 index 0000000000000000000000000000000000000000..699cd128ac317b02907709cbaa6069c9c049933f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/ctgan-n11-15215-20260521_074228.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4173a0beac1a376242e3fb93dc11207774d8adb81c7614c07168797f32c6694a +size 2886289 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/gen_20260521_074228.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/gen_20260521_074228.log new file mode 100644 index 0000000000000000000000000000000000000000..b0e9b74579f380326bcabb446b0191b4543ac640 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/gen_20260521_074228.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63596ec8b3aedbbc2ee5aa38e79e6001453b5389767c1a61199ff2a7a254efbf +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/models_350epochs/ctgan_350epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/models_350epochs/ctgan_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..62019c1a1046b861494382bb22b9f3462ec747e2 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/models_350epochs/ctgan_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3427afad154f066b730b131aa56c20491192a7da09b7d354b77e2c17416dbfbd +size 1039331 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/models_350epochs/train_20260521_065103.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/models_350epochs/train_20260521_065103.log new file mode 100644 index 0000000000000000000000000000000000000000..7eb1710feabe94503eaa5b77a992a18f7b241355 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/models_350epochs/train_20260521_065103.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b22fa7ad7c1a17b1155c35c41ff55720d299be91cd984fa402982134d9bafb6d +size 67466 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..dd569017d9b0b070603e9d16ac67bdc8f9757773 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed5f11eee3ef529fa3e4b50b40fd80d9a487344a67d487be81b841ee4f08ce5d +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..c951e2972f6dc6938c25a935bd7684d217a7a110 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:165265a019a5d837ba442e6d6508dd73988b6c7e6760d96aa92b3735aaa2ef2f +size 2471 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..7ec905c64220ee1eb5c23190ab375e6ab708e8c1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2514567fa317b8ab719cca956cd088c7fe6f2e3a5c5edb67458e78fbff80cc19 +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..e98e1d88e87e67184f7ba001a34e9c89961eed5c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17460c96bc7b714bd5290cc1821ca35206c1c700af0e2abfb8c93ea9d27e47f7 +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..92109556070714781916320193d6d0a078c29135 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea4a39444041f9023db44e798bb2fd4168ea84025ac042705fa0d2998217a07a +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_065103/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..06548ba463e92d8cd5c8de2bd611cc245579768a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_074345/models_350epochs/ctgan_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_074345/ctgan-n11-15215-20260521_083422.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_074345/ctgan-n11-15215-20260521_083422.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..7818374a4409f5beaf722ec2b0eba56a110dbc83 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_074345/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=64, + pac=1, + epochs=350, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_074345/models_350epochs/ctgan_350epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_074345/models_350epochs/ctgan_350epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..0f6225fe99239d79a36fa78846c7a0c0e74a0918 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f37bb728a9039c7eeb51488cf8b5f21c4dd3597e2e7cc12647def9b343053b31 +size 205 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/ctgan-n11-15215-20260521_083422.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/ctgan-n11-15215-20260521_083422.csv new file mode 100644 index 0000000000000000000000000000000000000000..75a90684dc755211143d4db2eb2e0e0a3ff7092b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/ctgan-n11-15215-20260521_083422.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0c57d2c94e2d0e41014fffbe5dd8cd790a3085e5858c33fa6b9cc629b8d08d8 +size 2879918 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/gen_20260521_083422.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/gen_20260521_083422.log new file mode 100644 index 0000000000000000000000000000000000000000..093743b7807b601ebe060290592ca93f23b73140 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/gen_20260521_083422.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:055eb051996a5d0e86522732f18619626221367490711722d40cbccc46d088c7 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/models_350epochs/ctgan_350epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/models_350epochs/ctgan_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..19df941577596d8298d82b498cf0bd5f997226e8 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/models_350epochs/ctgan_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c677507212f77f9fcaa7c6207724e808d022bbc6706f94f4eee3e487696327d +size 1039587 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/models_350epochs/train_20260521_074345.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/models_350epochs/train_20260521_074345.log new file mode 100644 index 0000000000000000000000000000000000000000..a4fbcbb4926eda5773b319f5938e91eab47ebb77 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/models_350epochs/train_20260521_074345.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c441330c8f9fd2e5d8556da494e88f956aed4d2fa125f276e63d8476027ae95f +size 67884 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..a21302192e954c27cedf4f5632cc35d48382963d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:533340b5e2856c2b74927ef6a8b8f8e5190998096a5f11c5e8f91775208062b8 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..1901f4b1fe47b023caa0d22739a46a68f24d0ffe --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6987056507d9941d0c7be5c34da59c72a8eee765685a2351966a94a9660222d1 +size 2455 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..9f99b873fbf6379915089eb24290aa11f0fac763 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3a2689f0a14e68c505e9194a44012f49d155d5bc61a875482ba18a4e37f2686 +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..53d393ae152a291fdf24c635c63bde2b6e9d9946 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f15b4b9d89d54b98e4602b680e243386295db199d38bc0e8796439f8d7e474 +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b8087785c871e8f3d7dfd6c1dac54fed636cc8bc --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c602604f8d345a94450d0995cbb821ca37576438c13b4f7a73d4a8a8fdcd68d8 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_074345/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..fbb0320621b2a8012eefe29d660423b079292837 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_083539/models_350epochs/ctgan_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_083539/ctgan-n11-15215-20260521_090147.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_083539/ctgan-n11-15215-20260521_090147.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2485233d82812d13232955112983fac5afbb9d23 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_083539/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=128, + pac=2, + epochs=350, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_083539/models_350epochs/ctgan_350epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_083539/models_350epochs/ctgan_350epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..f316f969f7c9b1b41f7ae873ca862f4c86900a0d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:731faf9493eb53f817202bfd405f1ef91197c925b54aa012407c2e7977474b3a +size 224 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/ctgan-n11-15215-20260521_090147.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/ctgan-n11-15215-20260521_090147.csv new file mode 100644 index 0000000000000000000000000000000000000000..89f923005adbf5ce36dcf1f75176d1d31030a9eb --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/ctgan-n11-15215-20260521_090147.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67494c8a78fdc1a3781797cd31542163a7dad5ff0d5b1ec8b822372c64423daf +size 2892845 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/gen_20260521_090147.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/gen_20260521_090147.log new file mode 100644 index 0000000000000000000000000000000000000000..19770cafc568949dfe64221121005391471c7e73 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/gen_20260521_090147.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ce0c5836ef39f4064b9396de8f61e7a2b598ce8ecb3e13bb7f5f2bc01269e9f +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/models_350epochs/ctgan_350epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/models_350epochs/ctgan_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..2b1c6c603093fce25ebf1a0303f84891cf5c993a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/models_350epochs/ctgan_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0f3b2390b0f9d3b08b5a0920adb917b25c3be6abd8610a56dbb138683fea0dd +size 1039523 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/models_350epochs/train_20260521_083539.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/models_350epochs/train_20260521_083539.log new file mode 100644 index 0000000000000000000000000000000000000000..9f7c6f143f1b9b9b36256d30139452379ebabc08 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/models_350epochs/train_20260521_083539.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c9f68a1eb881f19f3c698234b8868b6eafbac9303c5f0fccbfd0d35dddd070b +size 67857 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6ff09f458baf2dbdd83d8b874f30d3d748d1ea91 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0e3556455d10021660ad9be4ad40fb993948f8dac149eaf31b76b946bf12fca +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..57e6e746656d3aa111559c4b332b9c46fed3de10 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad0c10fa287b0b83aa43cb52548cce96bcbe14c10d58fe06cf9254e5b3981a9b +size 2464 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..43591d8c4d76ff57f1fa3b3a7f585241b2530631 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50a6d0b2c774cac12dc05cadf7a28aa02dfa15479c1d30b4fbc48b89b254e1b1 +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..91f18a37e97969fd08a9a6793862c75b6a9b52f6 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a91cbdf37659929492e85c4a1bf350c018011c717a413efc31171cd48d586435 +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ed206e2c06a28d5d08a60a219fbc238bab3357d3 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fe3ac1bfd7b49ff419c654f42c183d263815a9b896f385d8f1de3e5fe105837 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_083539/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..a6e19b0dc0c793ca9d4b4abd186120e6abf10fd2 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_090305/models_350epochs/ctgan_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_090305/ctgan-n11-15215-20260521_091603.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_090305/ctgan-n11-15215-20260521_091603.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..5d745c6b2d21683b82731a0d35fc2d7f84d15d30 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_090305/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=32, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=256, + pac=1, + epochs=350, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_090305/models_350epochs/ctgan_350epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_090305/models_350epochs/ctgan_350epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..8fd292da39c17b643dc941dd1795452d92a53ec7 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69bfd73d975ef3ae45b50f9266843d9b00b9387ebe8a0c1de9d6d14b6bce84af +size 242 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/ctgan-n11-15215-20260521_091603.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/ctgan-n11-15215-20260521_091603.csv new file mode 100644 index 0000000000000000000000000000000000000000..973c9fec1c29ab27491c0e638734529b4a61bb7a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/ctgan-n11-15215-20260521_091603.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24eeba087ea75d6f141b6e35497d686f58ad8fd78b95b6a28ff6ec0fd181808e +size 2897647 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/gen_20260521_091603.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/gen_20260521_091603.log new file mode 100644 index 0000000000000000000000000000000000000000..4d072338c3386dd36fe072e1103c368149061425 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/gen_20260521_091603.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3701dbe033373a40917da8edf89eb1e7489a0c5d4cf9b3b81fa3eb6039c9bf76 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/models_350epochs/ctgan_350epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/models_350epochs/ctgan_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..ce035e1ad2f8f566fca9233770738a94a40ccf96 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/models_350epochs/ctgan_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3e9f323606f4bf1a65fcbf04d2d82dcb6b51c146c61157aa24b7bd27894a4a7 +size 1079267 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/models_350epochs/train_20260521_090305.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/models_350epochs/train_20260521_090305.log new file mode 100644 index 0000000000000000000000000000000000000000..584fcb636674cfb92c3440d403922eeee2353dc3 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/models_350epochs/train_20260521_090305.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e533fff1843bdbc3fd9883c6b73aef07e5266aec58ec7f1f3b3236226096ef1 +size 67900 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c7e3a59bf98e3d975cfab3303d27b74c6b9af06d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa3bc7f08ff7d846b3f9c4b9632adb2aa7bd2babffadbe1e59c798faf04a54cb +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..92ab523989e2ef67022055b5a583ea3f5e2293a6 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c3942b3e5c4a518c11d453e3d1959fc91a6e2f173bae6c29433b1a1d523d03b +size 2472 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..d354a439fed0fedd609a4f0e304efa39470d965b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6f3320468ca64e331e5ec345bc71e8f8d6f01685a1273b600eeb4a060c348d3 +size 911 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..10958e3125d328758bebcf5f866c8e699f1b94da --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef9a4f8680c5b55a5a819480c1cf61316972748800e1efe95c5e463f9472316b +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..a0d676a3af400526b3e6d8e0701b6a6e4444afa4 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a91352a526c2363ad44e7c450828f838b2caef8d45ccded5d4ba48d678a9314d +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_090305/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..ac4c1a865ee2f53cf9d8fb34280b675dee1c8349 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_091747/models_250epochs/ctgan_250epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_091747/ctgan-n11-15215-20260521_094634.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_091747/ctgan-n11-15215-20260521_094634.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..d2dc3283708bb34e1686d06c56eb44076483356a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_091747/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=64, + pac=1, + epochs=250, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_091747/models_250epochs/ctgan_250epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_091747/models_250epochs/ctgan_250epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..ba9d67dbda1c2ba03fbeea2f000cbb3258bf4e8e --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c769efa18d5b59fd24239263a04ec9e1a767e7c29b06ba79d0f9dff56abc7525 +size 241 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/ctgan-n11-15215-20260521_094634.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/ctgan-n11-15215-20260521_094634.csv new file mode 100644 index 0000000000000000000000000000000000000000..26a10ac1a09c4c3c7fa2126892fd632c6903ced8 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/ctgan-n11-15215-20260521_094634.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3eb7a62d5af6067b7ad378229d3f23f1b9145da0de27d274e5ea9a7125ad55bb +size 2897524 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/gen_20260521_094634.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/gen_20260521_094634.log new file mode 100644 index 0000000000000000000000000000000000000000..9b4593a6ac4106430c8ef3f4d8bd9eca6835b8e7 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/gen_20260521_094634.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c08f9a5cf13834744ed9da2cd900e8235d2bf5161ece60dd2e9a4458dc2def5 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/models_250epochs/ctgan_250epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/models_250epochs/ctgan_250epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..d3d6a1750b327b0b545b7dd7a68798bdd0af7f4d --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/models_250epochs/ctgan_250epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d571c1c20d23ca755f04f8a84c123884669981885986847807bcd8eed72adfc5 +size 1036515 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/models_250epochs/train_20260521_091747.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/models_250epochs/train_20260521_091747.log new file mode 100644 index 0000000000000000000000000000000000000000..48cbf1de5fc919debf09d307b0974c9510c5473e --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/models_250epochs/train_20260521_091747.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a5ce1c32b1e3230ecb28fed7885091c7c94473812c2583deb4f5eb98e7a7a51 +size 49147 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc2a41e25d155139c335a130dac19de00d808fe7 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8597eebea783cafa291cc6ae9e023d62eb2921222a13a0d3c4b458443b3a91c8 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..e0e02b7616c49b944d56c9e23853c2526a0657a6 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc69a55e1597ee39531628684bb59540d240acec2f7cc847daf38664dc82fdb1 +size 2471 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..97c245a11b8ee2565355293edbfbfa3fe853cd49 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb77255fdb23550d6bf2cbb3293d47c3ffea423a5fa8f3a0d1713176d80f1e0e +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..235fd9d4a857825c1c1300c7d27466a4c388c0a6 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d60088150e4741cc77144bf3cec2aadb129ff38848a7ef2cfad1ad1646cebf8 +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c44de73e500653fd7c467fba2f3ff6b4631bc1e8 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39016374d78a75e8157b1e6a8ead36cbb593b97023eeff10700152949f4a0150 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_091747/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..d2f928b549ceae49f2fc9e68141c02b99a590e7c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_094751/models_300epochs/ctgan_300epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_094751/ctgan-n11-15215-20260521_095923.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_094751/ctgan-n11-15215-20260521_095923.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..dcba543f7ae18a7be685f77436f696bf506cd62b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_094751/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=256, + pac=1, + epochs=300, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_094751/models_300epochs/ctgan_300epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_094751/models_300epochs/ctgan_300epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..a673275de4be7f866c7285d3a2aabfe20b79161f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ae20264ccfea317197de78adee3cc9e12bdab07f4b1b6400c9d41813d57c36d +size 242 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/ctgan-n11-15215-20260521_095923.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/ctgan-n11-15215-20260521_095923.csv new file mode 100644 index 0000000000000000000000000000000000000000..e41bd038db5ab7f7f98d5437edd2a93c27c813bd --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/ctgan-n11-15215-20260521_095923.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcee1b20702ad931d400ceacb678d85b16948738e153577c12fbad363a61bd68 +size 2878607 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/gen_20260521_095923.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/gen_20260521_095923.log new file mode 100644 index 0000000000000000000000000000000000000000..08dd9b5ac8cfe53a8ffd2efa248b3e9ce51e8f93 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/gen_20260521_095923.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:895a6ba4319ede720a3ea5519aaa71b86fe5bc6c9a20e7ffa4b0880fc9bdf46a +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/models_300epochs/ctgan_300epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/models_300epochs/ctgan_300epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..39978c1bd5defff230012c387d377fd53eeaaefe --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/models_300epochs/ctgan_300epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68e28391c302bb4def8fbe8be862293b74f36718979d6f3236d76dcf236857a0 +size 1037923 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/models_300epochs/train_20260521_094751.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/models_300epochs/train_20260521_094751.log new file mode 100644 index 0000000000000000000000000000000000000000..a7bd6eeb19e19301dfe39eb5c6a109efb7096e53 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/models_300epochs/train_20260521_094751.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9a75df20b34288d1f47eae54bbf60407db21de43cd5d7d22fd1bb7cffc37348 +size 58542 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0caa368281137c408e5c57474284edd4c774c7fc --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d71fb46dffc612a130143c4fe7870aaa62811ccb18e954cb3be831b028574c4a +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9f26c99e51bf50997337751ff509d842a6946e22 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63b44c4ee3aeaaa050cc01545721e2131306a63df795daf4b0a4aff467a800db +size 2472 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..bf338fe06ac0112a807b88d68b9174fce787db91 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68dbf0b1810f5753ea1b7421e6c540bef859a5ebb9841fe3ed488410dafca4b3 +size 911 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..063b9a135e9cbef935c4d102c16c82e5b285f12c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:314ac073159e652d875dedeba28b3fea7e62b3d956e705f77ce5e7897d3f33b1 +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..cb2e29cda5c2bf6a7729fdd42127e3612722e14f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb716118fd7044f3284c4f6c194a59d6dbda7eb8b93762f6441b288d99369b17 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_094751/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..7229a70da9356a702c20d304376c114b4ba38b25 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_100044/models_350epochs/ctgan_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_100044/ctgan-n11-15215-20260521_105221.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_100044/ctgan-n11-15215-20260521_105221.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..c49091f91adb3e322cfd433a937973a2da93f176 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_100044/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=32, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=64, + pac=1, + epochs=350, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_100044/models_350epochs/ctgan_350epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_100044/models_350epochs/ctgan_350epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..243c6980f05543a87a2cfc4dcbef36890fb754b7 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8b884f96618ac49250e92f171b694d0287ce94b086f1683d294f7824c368568 +size 223 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/ctgan-n11-15215-20260521_105221.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/ctgan-n11-15215-20260521_105221.csv new file mode 100644 index 0000000000000000000000000000000000000000..e33f9a1e913e2a38b7f5c3dfbd2c0989137aad6f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/ctgan-n11-15215-20260521_105221.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8680c0107921c9a4cab2e108b3b9812fc52743c3df7a91acc2c490a0aef2969d +size 2888174 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/gen_20260521_105221.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/gen_20260521_105221.log new file mode 100644 index 0000000000000000000000000000000000000000..2d6220c50e90018ef55002dc3bd481718d3291ef --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/gen_20260521_105221.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17b517cb8016566de31dbce943385f2b1d1dea78a6e2289d27108f537f34c3cc +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/models_350epochs/ctgan_350epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/models_350epochs/ctgan_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..24bbf09e9372f89b40e8bbec77f7ce07198e4177 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/models_350epochs/ctgan_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5144491698bd3ac0e6a8bd11e3c086bb2cca865bcafd186f976ef999f6637b2 +size 1079267 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/models_350epochs/train_20260521_100044.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/models_350epochs/train_20260521_100044.log new file mode 100644 index 0000000000000000000000000000000000000000..d763b252012cc67d0a8aff07131dae3679727ce4 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/models_350epochs/train_20260521_100044.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31dabb99103cb79d654ae749a7fae7495a0e305fa1fb2223ea4db5e9d228f9d9 +size 67900 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..1496295c08b7b834d45fe4400731963d5af9b41a --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cb4f9b20f1c1e76560c3160a2950c5e9922fb78dd126d04fdc9db121af465a7 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..d6d7801d8da888757971c5f33992fe5436f9041e --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb14be78e1f3ca41e4272789e19de3ddb9f333f47695464b19fcae8eba58979a +size 2463 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..e0b417ed0ec80328f2ecd935cc9a5d130e2b2488 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:735652f451ef9d93a54303a2d14020fafa811f853d77b141f028804dae839f0a +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..670ecbf452da102cc2bf2e05dc18ed06300b9486 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:111506f7cc476c9d46972bb66d6eea4621b0d2cf34cdcbe6bd43959aad68f4ca +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c93ef03fb5fd8dd471ce2fe9d5f5e1e8c18a6a6c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8820c0e9bed753a05ad7650e839253ac36c43e04ea0327b7ce29cbba23d4a70 +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_100044/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/_ctgan_generate.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/_ctgan_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..bfe4f2c0e77d98141338e46fad81a4f536d846d8 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/_ctgan_generate.py @@ -0,0 +1,18 @@ +import sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN +model = CTGAN.load("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_105338/models_250epochs/ctgan_250epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +sampled = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +sampled.to_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_105338/ctgan-n11-15215-20260521_111238.csv", index=False) +print("[CTGAN] Generated", total, "rows in", len(parts), "chunks ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_105338/ctgan-n11-15215-20260521_111238.csv") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/_ctgan_train.py b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/_ctgan_train.py new file mode 100644 index 0000000000000000000000000000000000000000..53ba64d0e3cbc4546b5f230aa8bfdbc7690ff953 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/_ctgan_train.py @@ -0,0 +1,17 @@ +import pandas as pd +from ctgan.synthesizers.ctgan import CTGAN + +data = pd.read_csv("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_105338/staged/public/train.csv") +discrete_columns = ['g'] +model = CTGAN( + embedding_dim=16, + generator_dim=(256, 256), + discriminator_dim=(256, 256), + batch_size=128, + pac=1, + epochs=250, + verbose=True, +) +model.fit(data, discrete_columns) +model.save("/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_105338/models_250epochs/ctgan_250epochs.pt") +print("[CTGAN] Saved model ->", "/work/output-Benchmark-trainonly-v1/n11/ctgan/ctgan-n11-20260521_105338/models_250epochs/ctgan_250epochs.pt") \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/bo_combo.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..d9c50a1363c9c64a0478c09267f1ba5c6c661a4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4805759c5314fd11dd9f1994070e01bdd1921bc38e12ed83554c3dafc752a8f +size 224 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/ctgan-n11-15215-20260521_111238.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/ctgan-n11-15215-20260521_111238.csv new file mode 100644 index 0000000000000000000000000000000000000000..d956ae8b98c615a01d55b1cbbe123508a12af8be --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/ctgan-n11-15215-20260521_111238.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:330285f0897edbf9d48e48f6e2b73c512e7d9e931f3204d6850495eafdecaeb1 +size 2883440 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/ctgan_metadata.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/ctgan_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/ctgan_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/gen_20260521_111238.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/gen_20260521_111238.log new file mode 100644 index 0000000000000000000000000000000000000000..6c4afa98bc2610a1c2ded29a272b7c374f796605 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/gen_20260521_111238.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a7519fb14adcef9eb8df699211729bc991d55136057e5ce3e12cf095c0a3e42 +size 302 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/input_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..d34326e721f72e525965e85b659b2ebbbd364911 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f8b938efdb85281245fdd36e9bd0daedb563df2c8ab49b2a99cb37f1500fbf +size 1358 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/models_250epochs/ctgan_250epochs.pt b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/models_250epochs/ctgan_250epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..c6cbaa5db6f8fe2ebdec52302d135b26d81b1f21 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/models_250epochs/ctgan_250epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96a567e7df2934622dcaace491b51ed4f0645ab7c20795edc8e45dab2f6e3e9b +size 1036515 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/models_250epochs/train_20260521_105338.log b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/models_250epochs/train_20260521_105338.log new file mode 100644 index 0000000000000000000000000000000000000000..25c0a216cc3476b219e83c626527c000ce2ba20f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/models_250epochs/train_20260521_105338.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a126434ffe09355c7a6809327e10ad4ed512d094d4c12b12ce2ad08a3140edbd +size 49147 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c5e6d2ff75b5c7af374df95b740f1c461ab66ea6 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b3e6813d6c31d62332b3441ad8fb581796f83d05d160782048ddd3039cbedf1 +size 6079 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/run_config.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..ce753a63d28068dcaf2ffe1bff1216a03f7a3333 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f7a9d33335fbaa5f439b60de482d353e33122cc38d1c996f3fb9d8793f32947 +size 2464 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/runtime_result.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..6c9ac646350f68ef07aa7c1fd8f5cb13329eed7f --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9eeca342282d77f06deeea5ffa52fb27dec4472c4f3c1dbaea9144235506c72c +size 912 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/ctgan/adapter_report.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/ctgan/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..1139f80f0e997cfc6eb3e6656cb73cb63c766e7b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/ctgan/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1586fdd173966820cef4ff64b5b67ff4a34b8cb2886d83ff83d5919b8442c60c +size 317 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/ctgan/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/ctgan/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/ctgan/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/ctgan/model_input_manifest.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/ctgan/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..5b93690498b864ba5f9f336e5ff617d8178ef865 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/ctgan/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7de1b58a68c507cbd75325e68db5e1d4e8728615660313f4ab1b19e56b79f0f +size 6272 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/staged_features.json b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/test.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/train.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/val.csv b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/ctgan/runs/ctgan-n11-20260521_105338/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/final_recommendation.json b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/final_recommendation.json new file mode 100644 index 0000000000000000000000000000000000000000..ae292bdd35b0f04e43c8b905b2c74ba62ddd113a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/final_recommendation.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce13cccde3e4862421f37a02398652d6f093d55cae06e5c8d95215d054514f31 +size 2868 diff --git a/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase1_manifest.json b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase1_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..89bf0c03503eb4288b286f46435c353f652fedb5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase1_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e74f7453c284248f8333f060a4176b43d845d8d856aef324bcabeac84e65ed0 +size 2233 diff --git a/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase1_scores.csv b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase1_scores.csv new file mode 100644 index 0000000000000000000000000000000000000000..9970f543052b0b0964003f3b9f42f5ef35b55551 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase1_scores.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62631bb794709e97770345e66a18009bf5b95dc137db239ca20a9112efaf4367 +size 3226 diff --git a/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase1_summary.json b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase1_summary.json new file mode 100644 index 0000000000000000000000000000000000000000..8470f39fe56e7c13c8b2fa663092830dbfff2e1f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase1_summary.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7878b116f1b696c83bf494e81580723c9c02e62be13478dfa0d84e57856e361 +size 577 diff --git a/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_manifest.json b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ae5c81a1d055209fca758877cb6b8793ab45c967 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58f00b0975d6962cb9ae71366e866698dfca850606b2c87467afa4ef17b9a5de +size 2064 diff --git a/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_report.csv b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_report.csv new file mode 100644 index 0000000000000000000000000000000000000000..f1ce32980750ea095ce4086dd6c72ba5f0adad18 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_report.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14ea425d8a49797acd6cd491fc803abe79bcbfa0977b0efd2df4164f392d8436 +size 901 diff --git a/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_report.json b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_report.json new file mode 100644 index 0000000000000000000000000000000000000000..c9808e05e504db4e4c2dc5f2cb80a30a2c7f4a37 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71f99c5858805fbea9de54638e571cdbebcd9d66b6658a1b9be8f10bd9cd67df +size 2979 diff --git a/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_scores.csv b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_scores.csv new file mode 100644 index 0000000000000000000000000000000000000000..124c37043827add8bcac73962c8d97918675c716 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/phase2_scores.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:891b5782d33dc1f43c85c9a6a32d81470b64e6323f235c116643829cd33021d8 +size 2716 diff --git a/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/sample_manifest.json b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/sample_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..899d2c78dfcc4e296d408d2192293216b40bd7c4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/_rf_tuning/sample_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ad58b7bef2caaac4ba9c2c79a56f1a2c4adde5d99d767a36bc3f0b4a3265cac +size 637 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..65bc7d9ec81e8148e0d5170a6fc8847373d79014 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234105/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_150.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234105/tabdiff-n11-15215-20260520_234446.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..6334ae70ca4b5433d249aec08fad4e66a7fc5bf6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234105/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "150" +os.environ["TABDIFF_STEPS"] = "150" +os.environ["TABDIFF_BATCH_SIZE"] = "512" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "512" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "100" +os.environ["TABDIFF_TIMESTEPS"] = "100" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..f4e6796bf2804fe73419f9e17e8adc8c19093680 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07dce615846a91fd12d9865e7bbb33208c78ab2ba1e571f70b672a2a63c4c3e1 +size 125 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/gen_20260520_234446.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/gen_20260520_234446.log new file mode 100644 index 0000000000000000000000000000000000000000..d358cfee553d9c1e9147bae06c8662695871afeb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/gen_20260520_234446.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef8976c157be62297b496e5b7ebc170646d961806b1ed6d3096752626e961123 +size 30835 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6c2f0ef9031836ea1a0d4a45b9531a083c49717a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d97a484ba138fddc191525734bb29de35d88c6b961e9537867540898fe083f3 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..ae7e1f97d3ffa119a5420efb8c3ba21fdee83c19 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1e9de8692a4e3b341d247780761af5600089fa0512cbb32db8889a10ebc834e +size 2429 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..7833cb8422bac09d2c22208a46a75b6593c2e4e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17f85b0661755c034a1c1280f2cb48bc0f1ff992fbdc2ac95586d4436f62024a +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..21e4a8cb05a92ddbd4acbd84603e822809bce327 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c4b97feb1f26cfe43aedba6c1d6ffba967514ac8dec964914791ae518a3a188 +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8bc65344afb629182ae572e718c7cb7457056ed7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4a2f6c22b20e52a62e8b206a2d4a4c1e6c1930b61622347d628ea5fbb2134df +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabdiff-n11-15215-20260520_234446.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabdiff-n11-15215-20260520_234446.csv new file mode 100644 index 0000000000000000000000000000000000000000..6c886f59dcb6e2f123e0fdf9d1e06afd8044ccb3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabdiff-n11-15215-20260520_234446.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42fabbe79d7cfab8a422eb58871bdad5511e99b9cff4f5c0aad1556bbc34f6a4 +size 1558391 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..aba2334f967a30f3741bdd5aeb5ea130e9bd50e2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71b01ea78f1ee611ede02e13d3aa4fa4b52bbdac11632e91dc91628df2e4effc +size 534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/train_20260520_234106.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/train_20260520_234106.log new file mode 100644 index 0000000000000000000000000000000000000000..fe9cd885db6759122e704f64f4f99481c81cf41f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234105/train_20260520_234106.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51f85002942cfaf85e8c2d7dacf4e895e0381312756644045eb021c95e396d6c +size 973346 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..f0ed979c380edeac7ced37211b8e4ab6390fc91b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234635/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_300.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234635/tabdiff-n11-15215-20260520_235902.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ea62ced94a421cb33f0cb11ff0ee8e5ae32dcc8c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260520_234635/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "300" +os.environ["TABDIFF_STEPS"] = "300" +os.environ["TABDIFF_BATCH_SIZE"] = "256" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "256" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "256" +os.environ["TABDIFF_LR"] = "0.0002" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0002" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "100" +os.environ["TABDIFF_TIMESTEPS"] = "100" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..de4d4a4047de6a6da464e140f3b8e056fc198959 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f674009e181a29264c8b71d90009a6208a22987878140a604b84fab6f0e5456 +size 126 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/gen_20260520_235902.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/gen_20260520_235902.log new file mode 100644 index 0000000000000000000000000000000000000000..4fc3f50d9f5ac5406b909af183e217fb19a30938 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/gen_20260520_235902.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7edfecfeae25a38758210cbe91f4ad3fd792551dd6d41b00610f8fcb3b5ba58 +size 51794 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..91aaf041cdadb9d3407ec8035cd0b04dd21a38c2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28f14e1d3db048a54d88ce54906c32028fe29f7f643c31cb6f1af463bc0b5bf4 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..3745385ee15e281926e4453fd11ca220cd3b369d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c151a2dc10d280d165ae5f6450a66a82b439539f6ff72fa860f163bab2c115f4 +size 2430 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..23431ce1012251cbe9b547222724f39033614611 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d8e0f9eab844781d48f83e3e68ccf90ea38ec7f58fd9c4b91fbdca9d36a41bd +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..2406b2e1afc85ab9fd9d1ab0f76add958cc5fd27 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e83cdcc343332427ce04079ef9107144aca548ab236052aee5ddeb0bc328b775 +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..5ba0708e889eca5153f4653676b465371acc4f52 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:542a14975361d6cf0e1e1463396c9897b0daa4c6cd825f4be197d387b4083802 +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabdiff-n11-15215-20260520_235902.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabdiff-n11-15215-20260520_235902.csv new file mode 100644 index 0000000000000000000000000000000000000000..c9b13de2f2d64a84b538c2d9bd43b882147ec37f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabdiff-n11-15215-20260520_235902.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2677c99e3a1c62dd5116f03aa1d553ee05c5ac81131add176d7b36cecf67538 +size 1553820 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..e2e7b6fedf7b6fb83e5edd0d62d065cb57a4f745 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5c66267009cef067264fd148ec21111d509a641457d2242276d59335e21a9b5 +size 535 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/train_20260520_234635.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/train_20260520_234635.log new file mode 100644 index 0000000000000000000000000000000000000000..3100e53997306a99ebe60c122f49bcb8930e2021 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260520_234635/train_20260520_234635.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cefadb72cab0ba9f46d894fa7e1473b34e58cb98411230000f870e789f8f26a9 +size 3778033 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..79aea83dd48f223bc1b5b6c6b6de3d0f643d820e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_000109/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_800.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_000109/tabdiff-n11-15215-20260521_001617.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..9440fd3e707aa4299687aceff01a3cc34507189a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_000109/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "800" +os.environ["TABDIFF_STEPS"] = "800" +os.environ["TABDIFF_BATCH_SIZE"] = "512" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "512" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "50" +os.environ["TABDIFF_TIMESTEPS"] = "50" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..020e1b8ab617e8754ad01e8f7a79a25d40d8f7f9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54ca6d84b136cec5c5a355b5c56d56c4b8a1a9b0217f653a0045a98b11cf717b +size 124 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/gen_20260521_001617.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/gen_20260521_001617.log new file mode 100644 index 0000000000000000000000000000000000000000..1a21e493d5fe71f84ada37f0b16789a2db9ebbac --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/gen_20260521_001617.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec6553274d217f992775ab73510d157d69cc52391df62fedd7eec2856578ddb5 +size 18328 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..f435a5c9446364cff2c8ebddc8aeb522c73b99b0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:168560fd79106aa1cf17bd665df8e40f2eb09ba953f1ada777b303d93bc0c066 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..4f501488d13fb5b8d1a90f8a37b8111878132545 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:012912cf165214fc3bcedc45248302626b68b963baa95f3ea8b3b2f346376ff6 +size 2428 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..4c917bc8d4584029d7649c23de0e9c5a5b30855e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84be5a970d0ed792c1661aa81b5e5029dc3f113da8dfb8aa2e1d661e1a1e6ee3 +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..596d2baa4a321f4bf02b0f99e8d235134b52fa70 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcdc431e6166214cb8425ca203d58acb3e3cf6dd36bd9eb1f7da02e2c891f31e +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..916933d2b255b8716d43eeb3409ef5185ba03f96 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0b176156c5e0e8c5139f0a783dca67354e1d6afc9e5d11ae0bb68e69f56d4c8 +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabdiff-n11-15215-20260521_001617.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabdiff-n11-15215-20260521_001617.csv new file mode 100644 index 0000000000000000000000000000000000000000..53712757f09bb0d8b90e2fa99ece5a3df09c5c14 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabdiff-n11-15215-20260521_001617.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40ebe226bf3f1aa36f3304a2a188512e360c0c7f4c1f34d57b48ffa7c132d259 +size 1557038 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..2c0bdcdfcced82934a614e8abcbf3b98a9ed7c72 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fe7bee8f03564ee7bd180cc0926feb46ef79a3a81f8f74a0a14130a8230b079 +size 533 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/train_20260521_000109.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/train_20260521_000109.log new file mode 100644 index 0000000000000000000000000000000000000000..cba0a6e83885668508bc63977d7e0c376ef8f1bf --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_000109/train_20260521_000109.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80fb7ff4355048ff31cb6e97468bb3d96b4b8b21c37d625a8d98dedf77bfa493 +size 5014543 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..b3c183adf7044150e827eb435e31cbf31cf67de0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_001748/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_800.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_001748/tabdiff-n11-15215-20260521_011559.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..396de9d9e83f394cdcc1104b2873c228b1069f21 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_001748/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "800" +os.environ["TABDIFF_STEPS"] = "800" +os.environ["TABDIFF_BATCH_SIZE"] = "128" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "128" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "128" +os.environ["TABDIFF_LR"] = "0.0002" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0002" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "50" +os.environ["TABDIFF_TIMESTEPS"] = "50" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..f7b97373210c7ebde9f4cbb27173a99c89e58992 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1e527c74b07d856925d4fe949d2dba336b858554dfacaba8671894649807e64 +size 125 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/gen_20260521_011559.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/gen_20260521_011559.log new file mode 100644 index 0000000000000000000000000000000000000000..3668aeeed0fe1ab1d17ad159f995d814f519db03 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/gen_20260521_011559.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:375925b8fa3b55ca6dcc55c68942f07e98905a50fded2808d9cda67078bcd152 +size 59668 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..968ed1e0ecbffc48e1bc6074bd6440dc0e6c5641 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c32bca6742778e3aac2b2fb6b8658f7e815b3c9696da7657905caa8598e2cb6e +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..d069e19815a5f6916df5960087de34fd891e3bad --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5b10ff39d82b28acf5c92de98a4c21fa99ad85f4a79e5ee1e4ee174f541f73c +size 2429 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..6b01a005f5d03a2de15783c4fa1e7da027a24323 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:775da0484d394b2484dcf8ae764640ec5ee14c6671c87283704acae6b31f2462 +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..ab87664cee7a13fe8647f853596ee6c0f5c1e52b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d73adc4b1902201bc6f75039eb6d0629be42c53aff4745e6e7a5de1bc2e2018 +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..60edf2b25ad45caedbc63d7eccc2a3f777955878 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9ed66159633a9cab7638a161fe418ebe9e162ba58fc5985feb39cbb98b7e002 +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabdiff-n11-15215-20260521_011559.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabdiff-n11-15215-20260521_011559.csv new file mode 100644 index 0000000000000000000000000000000000000000..ff7cfe623a962ba337ee978d0c9e9b01c9725af5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabdiff-n11-15215-20260521_011559.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a9e23994643b240d4c36c9e1abf18ceb276927b57e88d5de3ac21f1ac34ecf7 +size 1555501 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..fcec88e20c2a918194bc4dce79757bc5c8a640fa --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:064568a7f796993d21670e8206ccbc7af891d6f74de2e0b6f0adfccb61124459 +size 534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/train_20260521_001749.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/train_20260521_001749.log new file mode 100644 index 0000000000000000000000000000000000000000..14d52bb1875e8b1d7b7d01d099166c538225f44e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_001748/train_20260521_001749.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cc0123e17157f9a8ad60fa7d1ccc1f5c11b003445b7c2011f95a3d65eac4875 +size 19744977 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..5efc7671262a6ee8341834012854278b55c0da28 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_011801/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_100.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_011801/tabdiff-n11-15215-20260521_011949.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..31d43bacbb529fa6a6a11d7ae5529ed59555ccc2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_011801/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "100" +os.environ["TABDIFF_STEPS"] = "100" +os.environ["TABDIFF_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.0001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "100" +os.environ["TABDIFF_TIMESTEPS"] = "100" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..672fa987cba31f2bff0c7e276a8458528eef7b18 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1400110f9049b823058fea5e5626590dee3ea569d43bf4c2223a2f4d1cbb2ea9 +size 127 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/gen_20260521_011949.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/gen_20260521_011949.log new file mode 100644 index 0000000000000000000000000000000000000000..4bf34966996597f8223178f5d406bc036c86b567 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/gen_20260521_011949.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad3cbf3e732edc1af4d8ad26e64dbc55cdf3c52527a17c19d1882540dca0e3ed +size 28071 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..639cb9c8f0e853518df1af3195db1d411f92bae1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bfdc60d875b549e4ab70ae2ec1b868f02d3236825e6e5f052f1de3d5b006438 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..74eea1c4f43c49cd3acfe402eb159772389ca7c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:974ee2e785c5b3f4b0fbe67499b485c792a03c3f9bb4b1aa46998ba9c826412e +size 2431 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..da31d2a25caa143ebbd79b950b6928aeaa32f458 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:426c93c87da0ae673841de6c58f2d4d8c4e06b81ea20af901abf0b7c84529646 +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..48713e9a1ed94d99337aa4038f2f407d21387563 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:172d65dbabb354fb9d5fbb1b0a2f95e40688546bceaf80b1ea8448846624c4be +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0d01022f4403afc66ccb3040bef020847294cd7b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfebbb570a74b17f4469dc936e7b7bf8d08003b2f5922d0ac7a17331ca039c5a +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabdiff-n11-15215-20260521_011949.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabdiff-n11-15215-20260521_011949.csv new file mode 100644 index 0000000000000000000000000000000000000000..4c5a78c60bde2e01e54489896b029da9a4f311dd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabdiff-n11-15215-20260521_011949.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2709f2203ce49795e466869c0349ae8abdff2a105447c652a2e14af681ec039b +size 1556741 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..7ca1bceb30a4b86740dbbf5141471493d4b40e56 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bd8a4e2f7b50ec335b33a9af2ed95576bd0f8626623e87d178e7e8162212790 +size 536 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/train_20260521_011802.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/train_20260521_011802.log new file mode 100644 index 0000000000000000000000000000000000000000..4e4d213c394c77aa2387915e38ce3ab35801f60a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_011801/train_20260521_011802.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e38277131cea5e2fef3b665bf39f051b00bebf949c854f7064ad7ea393a189df +size 356917 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..97431c744222986ea566e903db0716b49a3153cf --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_012131/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_1000.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_012131/tabdiff-n11-15215-20260521_013059.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..15007d4c44b22cb30caa7439c9193ef0755f52b7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_012131/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "1000" +os.environ["TABDIFF_STEPS"] = "1000" +os.environ["TABDIFF_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.0001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "25" +os.environ["TABDIFF_TIMESTEPS"] = "25" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..c26fc97a2316d5e7a7e6345f388724c8afc6d8c4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c9a3ca749570d38b05f995c0dad05c370fa99b9ab97961ed19df8e1a444036d +size 127 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/gen_20260521_013059.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/gen_20260521_013059.log new file mode 100644 index 0000000000000000000000000000000000000000..92335068a67e536927cd764bd8a507215376c31f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/gen_20260521_013059.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e088b176bb590dc1bfcba036454bf06e139e76c95acd824dce011d19bb190e67 +size 14035 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ecbc912109cc3a3fc82499993dd21dce81f0eac5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c8af817101bf1c3685de4d6c50cef1ec1f4722d8012096c7df1da74b7d77716 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..093fcb9e2100373080536a12d9d4be98ed07582b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:186b82505a9866462402302f9abeb75f862e40260aad9a8e5d283bdfd6846313 +size 2431 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..39ef4b5f9295fcab1416db5ed8dd5c32052a208c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc437f54bc69451c2bbb701614d5b75b8a5f488d649bc6d1ef68de3191618d8d +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..4b4d2a1796d17adb5b1ef81e31d62ad290f7c3bd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30867e59ed5eca6126070298e4f8d94c2a07f4fc7824c03b33663b5aa044b4cf +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6bb2c12050224add5dfbb642bc38461dc29337f6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf245c2fa01517e9ef985b89a7819a51f3d81bdafa2763cc69024ab6a809f5dc +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabdiff-n11-15215-20260521_013059.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabdiff-n11-15215-20260521_013059.csv new file mode 100644 index 0000000000000000000000000000000000000000..5e0f4043503ed73036f3a4c9bdbb82e646aeaac1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabdiff-n11-15215-20260521_013059.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:879503dec8b4f62b225cf6537b16862ae3a1d5b1795eb520438e02c33f38cc84 +size 1555313 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..68ea52e6b0c828b64ccd33935f81725c29dc9e9a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b00181745b981b667a155d1b014a71c6dcf54b75854195236ffda3f22cc9f921 +size 536 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/train_20260521_012132.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/train_20260521_012132.log new file mode 100644 index 0000000000000000000000000000000000000000..6c27c9a245e4b143d579b3d4fab127a7e5488df3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_012131/train_20260521_012132.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef11deed27d2c23bbd92426f466dd0644507affa3388c81d48301aae0a2b7be3 +size 3151622 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..8305c082e454ab56d40a119d7c887990077a67d8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_013244/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_500.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_013244/tabdiff-n11-15215-20260521_014217.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..8516f2f92882b2e00384a5fe2d8bac267de90577 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_013244/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "500" +os.environ["TABDIFF_STEPS"] = "500" +os.environ["TABDIFF_BATCH_SIZE"] = "512" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "512" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.0001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "100" +os.environ["TABDIFF_TIMESTEPS"] = "100" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..a710ca0ce37f1fd9ee809c753f19a6b91e8cf888 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4166274412b44c2429c83c5a9bb2c8aa190feb3cb429d3d939e8a32778761481 +size 126 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/gen_20260521_014217.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/gen_20260521_014217.log new file mode 100644 index 0000000000000000000000000000000000000000..79f514e81d8fe6dd8192043b63caf53745dcae53 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/gen_20260521_014217.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cf907db9db0868be3944887fe8c75f51388749d804c2481d2775e1f3b881bf7 +size 25773 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6f968b3b9a7d20a1631ce79ffcc54214b1c48368 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e22ea7015fe14244cae4e164d591a39101ff4da703a18f0c215351bc05a8c73e +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..c7602562a5eb10e9309b1dcda5dd05feb6279a8c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4eeec999065488e7e6601c1c6c599df5ba9d77079026d18a237e7b638dce5135 +size 2430 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..c3b73cdc56aebd255bce70faa34aa625da89a3ad --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f89217e1a87cb05ecb1c03cecd73938044f6d7798f2dbb034e4c7502d363184 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..fe1ffa35ca163d63a61062e4c60b631d8b6b3665 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47a5bc474958901896a3719feaddc00d1f4241bcf6d1e4468264be0b24cab68c +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d8471333dae2cb183176d8aacb4d7717d4c703e8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0295baddf7d081c316cb2d9ed9abe74f807d16da660911f30a6df75c42a759a4 +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabdiff-n11-15215-20260521_014217.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabdiff-n11-15215-20260521_014217.csv new file mode 100644 index 0000000000000000000000000000000000000000..da481bd3db965864c4f2b9b43e6df0604ec34b8d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabdiff-n11-15215-20260521_014217.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08134f224f9d14971d9a42d7921e0eac5efd75c65a9be00f38af935aa60a4a5b +size 1555881 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..5db1ab16a3e17b68d468036dd0e8de6971f15a66 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8e1e05512f275c2c367d46baeed3a89c32bff99e0e3111921ab09400ef9c6eb +size 535 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/train_20260521_013245.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/train_20260521_013245.log new file mode 100644 index 0000000000000000000000000000000000000000..2603eb105c545f83de75ad1f9e5fb39d61b6a424 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_013244/train_20260521_013245.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a61fef540ee29f72b2a2902ffa3df0185a8ee31ad0ae3badc40758a23b3e7627 +size 3130516 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..1fbe90daaabce399b54d90b08131fdf997ef7a1d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_014400/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_1000.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_014400/tabdiff-n11-15215-20260521_015415.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..74efe7649a1781290907277a7d5885922b61eaa2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_014400/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "1000" +os.environ["TABDIFF_STEPS"] = "1000" +os.environ["TABDIFF_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "100" +os.environ["TABDIFF_TIMESTEPS"] = "100" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..a674b093d594220f9af72e906270ab76887480ac --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b9d070cb4e4b82d2b12ef24154176722703d1bc3e8fd9d58d161ed56a6ebde4 +size 127 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/gen_20260521_015415.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/gen_20260521_015415.log new file mode 100644 index 0000000000000000000000000000000000000000..ce411049c78fd3eb054f72fa38357248cf49e1f2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/gen_20260521_015415.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c21e241ab2acb249bd95a5bc9cce1942f574e58ef18ebe19d8dd6cf4f13358f +size 25478 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..54228e6de88c1a17e071cc674322db3f5a44edb3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f458d29df35c3d0bc0d8b8d16044ce408b6203e18ff92b4a58f14fe240b8f5 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..84666b3101f98d70ff001fe2118237d3b81e9f3a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:046bfe30d4f454cae221ed563bf8d2fbf153d49d06f8a6924f3992278ce860c7 +size 2431 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..31300998b06f85c83760c61d9a1f320493c220ab --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e4f80d8ce89f3a566bac515e486e274251d6b6d2a6d09f906a126cebec31212 +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..837945a80c7401561ba2e38f25d554507db2a3e3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b9e94b106154c40f9f3c55d01b4d9b2fe5a7dec0265456afd0759a147512581 +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..00213783b97dd401340ccad52226da001417d8f5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7731f5508f233db91c542479035845d85d560d7429cfe04a319ba45cd65320bc +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabdiff-n11-15215-20260521_015415.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabdiff-n11-15215-20260521_015415.csv new file mode 100644 index 0000000000000000000000000000000000000000..3d10c32c6979241bd139b27628d7af1dd7c26e2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabdiff-n11-15215-20260521_015415.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee4b9a029578add5cc95c8e4ff42b680517c77d9e9d80005f58894f60e22f9c0 +size 1555583 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..99a5e38c3dce9171bb8c7d342077c61c28d31f60 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d2cdaf0b2e3c0120730479e238674995eb197a66b6afa2808fee1070feaf3ae +size 536 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/train_20260521_014400.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/train_20260521_014400.log new file mode 100644 index 0000000000000000000000000000000000000000..b6c1952854caaea29d4d885966f99476598232c3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_014400/train_20260521_014400.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7cc12dc633453582adb749323cd35fb93aa4cd93448cafa80434089242bb927e +size 3195545 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..741943bedc5b8482eb674c572b5a8425f0919b4f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_015556/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_800.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_015556/tabdiff-n11-15215-20260521_021045.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..1699be873edd560efcfd3d298a1d92fc979f92fc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_015556/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "800" +os.environ["TABDIFF_STEPS"] = "800" +os.environ["TABDIFF_BATCH_SIZE"] = "512" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "512" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.0002" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0002" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "50" +os.environ["TABDIFF_TIMESTEPS"] = "50" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..2688ffcf5e7e5f3039d8f562edc434275b12bf58 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3a6ed32de28cdd64c5f7a6e0ed62974333462d9dcd28920da4ada18d6dbd131 +size 125 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/gen_20260521_021045.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/gen_20260521_021045.log new file mode 100644 index 0000000000000000000000000000000000000000..715bda711a31e0d8543ea03856c091f7ac8daf7b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/gen_20260521_021045.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac21495a65e9908eb76b53c276c6113d30280307450c2d34f5a682e7949d4d5c +size 20385 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d030324f982b3c29510c99206e393798d649f279 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0b7e616ad8c8fd9ad41c7ea28e48e4b3568578267d76dac45a0384e9b69e53e +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..d429178041e22369ed79c8a355a4d245afd702e5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e918d167fbe7c9a8169245dd88359a79676a483be8b61dfbfcc0a039fe5710b +size 2429 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..1d470b4008f2fbb97455f24ce066d59d5789b5a0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfb2fbfe257c57a3abc2747b927944a2a440e5bbc691cd2cb454d6fa547a2b24 +size 914 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..52b5463a266204636e663881fbd1f150340126de --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b452c504bb5f216c265b0d52c85c371d9c91e752352f7934db158194b3571f0b +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e35140c85ed191ea7924e681df5994891f280812 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08769f02c3a73b748ae9da2394bda0f870fbd91264645d5dfc1d21f965cd8040 +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabdiff-n11-15215-20260521_021045.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabdiff-n11-15215-20260521_021045.csv new file mode 100644 index 0000000000000000000000000000000000000000..1cd74cd88d1bb3a4a7356b0997e788f969fabeb0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabdiff-n11-15215-20260521_021045.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3284e3f55866692fd547a7c99d627d40dcc6a70f5267b61b28c3f32d269dd4a +size 1556301 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..495474b9475110bba1c28fe50856238e8a117f34 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5940cc26d319320d3b0495d9892cb3e2ef7bd0ede97afc861b165717067d21e +size 534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/train_20260521_015556.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/train_20260521_015556.log new file mode 100644 index 0000000000000000000000000000000000000000..de2de6663aac10e90c415f363b265b4767593afa --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_015556/train_20260521_015556.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26f78c55eb9d5daadc8f92bbf5a819b6899e49d599e3dcc14448f9db45a477b5 +size 5015131 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..11bc2c39bcd67336c6f0d3ce0084a7f1d9257df4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_021247/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_300.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_021247/tabdiff-n11-15215-20260521_023504.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..4de13ecd9ee88e97bde86769c7a71a4d8115f4a2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_021247/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "300" +os.environ["TABDIFF_STEPS"] = "300" +os.environ["TABDIFF_BATCH_SIZE"] = "128" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "128" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "128" +os.environ["TABDIFF_LR"] = "0.0001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "25" +os.environ["TABDIFF_TIMESTEPS"] = "25" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..cebdf8404e6e3f21231a7bbee767f30acf5a34d9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:772c37159198a3de2fb7544208992e3b57f871f88179d440544a28779d607666 +size 125 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/gen_20260521_023504.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/gen_20260521_023504.log new file mode 100644 index 0000000000000000000000000000000000000000..f4a5cc0fb2aaf15087717570936e4efb58e96f74 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/gen_20260521_023504.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9791c3628ab3a6877b9073826fbdb4a1ad87a3a7c7bdae0023d51cb358725ba6 +size 40003 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0500d7e86e9b63a3a53a5e294e0cb13d3f85761d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f458fca5f8e0e7055e549ea5e43b38be49e711c2a5b89c4ed02601577031570 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..3cfde87d9408ae0a645e7f7763ff3302df1066b7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:277561fedd63cd8b38b9e0066b2de448dc621b9c87f0ceda587e8943adc29ac9 +size 2429 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..8c5c6dab48b76acdd1683f17cbf9fced1cd33dbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efbdbc6a59594580733165c76d234b614de548ec5782437f125d68bf76f53f80 +size 917 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..87b9d14c7d6ffdec9c8d44994ac2ba88496579bc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1c48bf3800dd54f8ffa198457726481e57129fa9f7a94f6367e560595e52b54 +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..1e1f26743f33c3836476308772322e65e6b57abe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d5ed0e1454760e4b4dc07c6f98ccc1696a18e66a08dbc2fc49d104e80752d8b +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabdiff-n11-15215-20260521_023504.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabdiff-n11-15215-20260521_023504.csv new file mode 100644 index 0000000000000000000000000000000000000000..b9881693a536f6b99857d2ecb3aad10f7f9b4ab0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabdiff-n11-15215-20260521_023504.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0382d820b2583e68c09d160c20a33decc1455ec68b32c9f9c24f2df7d15218d3 +size 1557260 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..990527698328a99e897b7a4631916cced0ec28a5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2cb18e97a4a63b8817a418dcd92dc4e3a47212b2e3fb33a6d2ca6c767a28407 +size 534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/train_20260521_021248.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/train_20260521_021248.log new file mode 100644 index 0000000000000000000000000000000000000000..fb0537a76c18239c23179f7e827113b47fb69fa9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_021247/train_20260521_021248.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c186c17b1cfeafaff2ab3678bbc46e79ed2b72a9924e050e9ef8779e2e8db25e +size 7344023 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..bca5c5e8e7b00bfacb77dc95b98697559f6d02ca --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_023648/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_800.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_023648/tabdiff-n11-15215-20260521_030551.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2e4ffe149bbcdb6b7b8db5d680d0361e1bf6791b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_023648/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "800" +os.environ["TABDIFF_STEPS"] = "800" +os.environ["TABDIFF_BATCH_SIZE"] = "256" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "256" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "256" +os.environ["TABDIFF_LR"] = "0.0002" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0002" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "50" +os.environ["TABDIFF_TIMESTEPS"] = "50" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..0a46ae775736b09fb33a976deda004c6463e12b4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2711e98819f06d992862e9ea8e665eb5bc4fe7f48e60d628ebcd0c0d32e0a81 +size 125 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/gen_20260521_030551.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/gen_20260521_030551.log new file mode 100644 index 0000000000000000000000000000000000000000..9a4dc99f137e5834bdb86ac8320b52431abe3b73 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/gen_20260521_030551.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9ceecae826a741242d94fd90b3c475b195b5a75393655c5029c0e06427557e4 +size 36537 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..3ed372029ec2f4b702eb46185c569adfc2c9f826 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd404d74a83b09959debdbf77afbe06cf21df0c7cf24a7584a931933a00b2f23 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..7d68205242fdae1902b4ea58aaa2613cd5ffe721 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af8f90d4160e498677ed4a518267012678c1df0645b6d52fbe44e8ad15960b06 +size 2429 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..d7ed018ffd3d461db9248d94454750505dc6a04d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c90d028b1a4e9f7ebff35c2ab78075e63266fc0897a4e874f56eef6fd86c9195 +size 917 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..2a70812dc3d93731597a3079e83b33c349ae1a4f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:164278b858602d34a2c38d8572d9dd3a0ffb78b88a44e5fdc20aedf371579afe +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6390d16e009b82b3cf50224b7c93ecbb763c617d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0b146f5eebac94ccdfb9ef91f9a30517d36b34b16b3e711ea87a6b947b4c89f +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabdiff-n11-15215-20260521_030551.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabdiff-n11-15215-20260521_030551.csv new file mode 100644 index 0000000000000000000000000000000000000000..b135e3a1431bbff7dfe7d6e062d6b7b70cafbf3b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabdiff-n11-15215-20260521_030551.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a29ff2890f5f8c0f5af5cc9f06b181df9e3108e21e1b6e3be4dfa3d52972973 +size 1556511 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..41647f754fb4af19371118380d958d975885eae2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7fa3f3ac6130fe6b1153298d5c2a2d972e822714fd6467ee2c9aab95c21b7d9 +size 534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/train_20260521_023649.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/train_20260521_023649.log new file mode 100644 index 0000000000000000000000000000000000000000..8479de5a87afbb4cc2fdf1c48d96e85392cb0789 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_023648/train_20260521_023649.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc7fff243c841216f44bbbb892d35668c95e480c03cd35116d7459c25625a127 +size 9907758 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..ee20f6314eb4cbed0d58e5fab9d32b168dab8341 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_030737/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_500.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_030737/tabdiff-n11-15215-20260521_032528.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..f135d17aa4058923d7cc2bd4b41e0854c992a68d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_030737/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "500" +os.environ["TABDIFF_STEPS"] = "500" +os.environ["TABDIFF_BATCH_SIZE"] = "256" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "256" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "256" +os.environ["TABDIFF_LR"] = "0.001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "25" +os.environ["TABDIFF_TIMESTEPS"] = "25" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..7178061f418bb3af02ce8c33e8db0f7389bd3d76 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5964ebc47f3b8f49fc5740a9ece491ee9d104935e08be1df181b745f8dc71522 +size 124 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/gen_20260521_032528.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/gen_20260521_032528.log new file mode 100644 index 0000000000000000000000000000000000000000..ce0afdcc3348a954f86c47dac1121e9cbcd05f29 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/gen_20260521_032528.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34ca8a508ada3b0f3954986034356e20127ba8cce354fb302ce84bb2e307a7e2 +size 22200 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..506f00c664b0179f92352dbaf197c0945cfe5953 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9fa208c89601b8a4f5d7f4d486a99982061d41b189fdb32d60cc1b508f1dd12 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..94cc5b30e34b5e3f1cc6b2defc6c9a9ee8c0bcd0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5173b183a8dc50081cf44940c0ae8df4f53cbf2ad088cf9a5ace225eebf628a9 +size 2428 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..c406b4d0a213e9288cbf29ac9ef8e2e2a6010e20 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a8a8771a23ba123fda5100a1846e82f2be02f7e6a55d563355869da37b94e44 +size 917 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..fb66b16e2a336ac31ec77d19f168428bea57537f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34a2ac31a63a7517ef34f697a5dda714373f82f741438a86db03e5afe2690929 +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c32c62d16b62544b4747b42fdec151ce0c7b0b87 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75380e02092d2314b0a8d6cf0e9a16de910c690d94f20ca3c74c2aa846611df0 +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabdiff-n11-15215-20260521_032528.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabdiff-n11-15215-20260521_032528.csv new file mode 100644 index 0000000000000000000000000000000000000000..909ea3fd028a2ec20374eb0b8d6c88586870e7e1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabdiff-n11-15215-20260521_032528.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9cc978ce0f6dc85565982689d8b4aaee328ac1aa3cd688e1165e749e8decc1a +size 1556605 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..c598c57c5176247161c1651861e6ae249cc2fa82 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f22c0fe884308fa9e467057687a7399ab04ac3d2c09cf7df539389f2a4dc4229 +size 533 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/train_20260521_030737.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/train_20260521_030737.log new file mode 100644 index 0000000000000000000000000000000000000000..30f04f1f087c77982765a6ea6aca3866996f4912 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_030737/train_20260521_030737.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e8c1ed1be48d9274737e638714cdd25843647fc5cf1f50693447475e7e4d014 +size 6151486 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..fe242044fb009f28492db866ebf2d33f4d0f0d58 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_034608/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_500.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_034608/tabdiff-n11-15215-20260521_035620.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..f269289c38c66dd052a12c940fd7b9e199a6ba6a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_034608/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "500" +os.environ["TABDIFF_STEPS"] = "500" +os.environ["TABDIFF_BATCH_SIZE"] = "512" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "512" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "100" +os.environ["TABDIFF_TIMESTEPS"] = "100" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..626fde7e548bfd6f691bd21ef87bde8511e4c845 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51c0fd202567211cdd3d6e772f6e7e3147a8026aaeeeb03e57dc847354eb8f80 +size 125 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/gen_20260521_035620.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/gen_20260521_035620.log new file mode 100644 index 0000000000000000000000000000000000000000..bbeb2daa4df3716afdfd47afe90dc87e7f77adc3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/gen_20260521_035620.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a70108351ed31ee157b50c06b41157e3b182aad9b5289b348091c25006b486f2 +size 27789 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..47759ed14bfa8b25f903d4c6ec1b1f68529c2200 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a607680e535bdd5f927be69a0591e075946be7a7e8f71a1799fe4ec970a7e8a +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..063afc652fbc725da4c4db0500d3b1afd374defb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ec933cacb54d699145935753e3c711a45b01eb0637eba2bdcf4de92ff041126 +size 2433 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..331b8b8c98837b5428f3ad5389524e3e4a8f6d86 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed775bae5d69bfb7f7856971737f27a576f272ca8c67f8a2f0918a728fd20754 +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..0c8d32a8215e86b3995a353cd5f086d00c041dd3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:457a17f94d2ab11fcb561ac98b15d059f4c0bd7a6def0b5b912eb2120e60f187 +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..3dd2d9f02d1c505ea5e7d0812366a39b1db4ca02 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9a49d2f5a85c4d11b87c5a709c28614eb5be5f12f20b10cc14f70a37c121da6 +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabdiff-n11-15215-20260521_035620.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabdiff-n11-15215-20260521_035620.csv new file mode 100644 index 0000000000000000000000000000000000000000..b1942cc308e03432689c1b576e19b4d5d19ee91b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabdiff-n11-15215-20260521_035620.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9e1add8e8bd0a461d0f0227501f1f4f6f2a74108822c81e00d5d19993b2d49b +size 1555646 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..ed9bbc336a6c6f16e4ab6a7e1aced19fb945e4ab --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21f5a1ce12eb4326495daef46b965179f9a663c65fb09a69a51df8f304fd4067 +size 534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/train_20260521_034608.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/train_20260521_034608.log new file mode 100644 index 0000000000000000000000000000000000000000..62487bc2b89298a9eb1a6d5cb17c4a67cd787ed0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_034608/train_20260521_034608.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fa847736cc020e903e3329139b2d1d9ee3ba0acede7de863cf71a448f398e47 +size 3145284 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..37fde9511d7db26e8cadb7c8cfb3d7fb201df78b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_035804/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_500.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_035804/tabdiff-n11-15215-20260521_040729.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..6e1fec4e4e1fc0a953bd4819e75c8c16f304158f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_035804/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "500" +os.environ["TABDIFF_STEPS"] = "500" +os.environ["TABDIFF_BATCH_SIZE"] = "512" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "512" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.0001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "25" +os.environ["TABDIFF_TIMESTEPS"] = "25" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..213910deb458f390ad56eee00956b075069fd1ca --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e56715bbf6e0cd4990b3e0d2b5b783d56afedb66bcb460b2d03f96869f47db1 +size 125 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/gen_20260521_040729.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/gen_20260521_040729.log new file mode 100644 index 0000000000000000000000000000000000000000..78e9099c5b37129cbc57c1dd011712675c0de756 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/gen_20260521_040729.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36bcc05f889fca11bc6b43cee520ca9ac4821db94ec465de1873a83f7c5874f3 +size 13069 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e221a1aa45452714216407bd019005126b393ca6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2eb47ae72c0d6acc47a93b1c5f633ccab874e689ba024bc8ac52a309f327bd7 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..a469053faeff727bb3aad13f8776ea18b95ab002 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6c577a090c6e4594cba6aec7a31dc93f50ec1b9bd84e5b4e36692a3ef22615a +size 2433 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..933c45fdd244bdb08bc397f40ed336cc59693062 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98b245b4bc8a855caafbe1a97b721c5297187c029e257c539868359764a24d5f +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..e6311460afbe474e0598df9e47ff9eeff30debf2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da50dc27517dc399660733be2f25d0dbe6d06c7783faa512dabf97322dc9649b +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9bf6c7aafac82562eda3bbad464571b7cd33cc30 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6688cb9bcc21c08148b2f89f42cf34a950323b2c23a7f4ca9e44473f1a8487d +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabdiff-n11-15215-20260521_040729.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabdiff-n11-15215-20260521_040729.csv new file mode 100644 index 0000000000000000000000000000000000000000..36c1f2d0426dd34547b7b609d2bb3351fd73b2e7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabdiff-n11-15215-20260521_040729.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d058fcfb199e48246485d996ae3d7376903c2aad72e41a6c0a7da0e86782ef11 +size 1554244 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..0a7fcd0edd445df7018d0378fd16689755e1d690 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1e792c57347b045635e0602f9573610acdbf6cdc7ac6c94039a57503ef3cb12 +size 534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/train_20260521_035804.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/train_20260521_035804.log new file mode 100644 index 0000000000000000000000000000000000000000..77574ddc62bf1a0c290ef7d6ce83b2fef68ddee2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_035804/train_20260521_035804.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3aec034ed7c63ee59cdb0b019837dc5aa3985f73918113d9fe1c1ae648564b06 +size 3114503 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..fe844f585bfd83dab86cd22387c2f7fae3fd6a36 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_040854/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_500.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_040854/tabdiff-n11-15215-20260521_041828.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ed7f45e791bb17ed7e05f599518a0f2ab6e3f9d6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_040854/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "500" +os.environ["TABDIFF_STEPS"] = "500" +os.environ["TABDIFF_BATCH_SIZE"] = "512" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "512" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.0002" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0002" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "50" +os.environ["TABDIFF_TIMESTEPS"] = "50" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..ef189cdc0c9ca65cca374ac035bf04c85a4be067 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c224f4a9ffbccf7281e54407b49644755f4725bbea981213183a41cf19d7d38 +size 125 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/gen_20260521_041828.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/gen_20260521_041828.log new file mode 100644 index 0000000000000000000000000000000000000000..12d0caf685ddf897e38d2d6d6165a9ec772dc8b8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/gen_20260521_041828.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c9f629918bf543a9f776725fefb04339033522051cbab68e81fb456af86fb73 +size 19740 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8bc75e22a01c4accefb4780b2fad1ffcc70cd630 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce48fcd83d3d2a3aa23ea4180c3e2170c8bfbb8a5a54b9a46059a7b958de7f13 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..8a034fc6a4347cdc28e72e9a695f702ef116a6b2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:924827df2d3c4b5816114e8849d714d534797b0c6ba3d9738d7a14afa6646dd0 +size 2433 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..44aa6c8c3d40bd3885b4340c2ad036510a34ea18 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3b209792467b84ff4a040d731cc6508669effcc7c298f4aa737e67c0024ef01 +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..02fd60fd75cce241891786096fa58374845f730e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d920329e27012045407be525173ea8c4b0a12c67b431a995ed158bd96c36dd08 +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4000a4240791699ced2e0f171b86e35b225a169a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cb4607115cadc6b7ce71b752376a199f93630beb8745c36dd1e9deef3dddbe2 +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabdiff-n11-15215-20260521_041828.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabdiff-n11-15215-20260521_041828.csv new file mode 100644 index 0000000000000000000000000000000000000000..9ee7bb6c76b6436d008ed85b1dd83a7869625133 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabdiff-n11-15215-20260521_041828.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:509a16c9bf9ae29c34ae7fa9802c06faf23cbc514ccf04b3316ceec0541e9be2 +size 1556729 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..8833b3cd0c17bc38075ac13160de782c305ee8b0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5e1e8a185054087426b26b74cf3d58c6a630bc56a5b80619312164b60b3ba24 +size 534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/train_20260521_040854.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/train_20260521_040854.log new file mode 100644 index 0000000000000000000000000000000000000000..3197c8b17885537c7fea60a057f7bb2ff10d44af --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_040854/train_20260521_040854.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa99304ece97e10ac68f453fa2dfde43a135edc1478438be8279ff9823684332 +size 3137160 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..439cd48a041682553f6e4b644353a36621866835 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042002/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_500.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042002/tabdiff-n11-15215-20260521_042511.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..b953fab112e00767ff9a751fc3a9a269be473b81 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042002/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "500" +os.environ["TABDIFF_STEPS"] = "500" +os.environ["TABDIFF_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "25" +os.environ["TABDIFF_TIMESTEPS"] = "25" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..166415aed8a0d131ee6bb781878572c416dd9851 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fd216503d727d59fd10c907b7864e8ad21fa25d76896783c679a0a402ef0e1a +size 125 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/gen_20260521_042511.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/gen_20260521_042511.log new file mode 100644 index 0000000000000000000000000000000000000000..30dd8dbea9c2f2ced1471b62547dd9a2fbd2af79 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/gen_20260521_042511.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:337278cdc83f1a4e44e41e768b175322bc599e1bff9576a8d952cb0c73fdef23 +size 15535 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4c3fcd9ec7bd400e911a3b1c371156364b24c8cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3fbb2024e02e127cb82985cca5a0e35567e443d8fa38308b9c21a53d2236249 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..70ab1b5c4f736ce092775353a5a4af438f41c56a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beee7ea9b1eb4a46dc917f34f8717faa0208282183e5274692a7b8de325a6ce1 +size 2433 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..56274454ef0832aabc1fd66b3224eec395d8003c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f2a86e2c1189b08f15eaa8a4d4ce77026aab07a2e776c2862704f7255cb3c54 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..3162d44762aa4375bb52f3b11e331cbd0d79dd7f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52b16bb6bf9c2a5ffbd901f5060e9742bd7121720c5381fa766cc5228ebab06d +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..40dc4ad815a6daedb1d965e4d2006bcd2cea7532 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df4b2c25cea8833121ae870ab9e08ca81025ccea2785387ed95c0f9acdf8792a +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabdiff-n11-15215-20260521_042511.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabdiff-n11-15215-20260521_042511.csv new file mode 100644 index 0000000000000000000000000000000000000000..b491211542a839f2b2f69b2f7c29a595ac18e721 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabdiff-n11-15215-20260521_042511.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:550a44afe39e67a534e0429c9325d644e41f7741f034a6edfdf0ba9fb00803d8 +size 1557516 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..4ee2a0f5cb40b8139336251c6e32ea704430a167 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc4b21631983e9a08e5c5d310311b1e568292bbd7140b800fa8ee3cc63860c68 +size 534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/train_20260521_042002.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/train_20260521_042002.log new file mode 100644 index 0000000000000000000000000000000000000000..42002ac2cc1323f8de957ca2c9d43e9582b4b271 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042002/train_20260521_042002.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd6cdf34fb01d07886f07fb134f809b8fc37e6f94e811842247ebef3a187f614 +size 1582017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..2dd63877b8bc471f40763a3366f392fdcd970bd0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042653/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_500.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042653/tabdiff-n11-15215-20260521_043250.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..158c7100aadec1b94bf3331d10abe4e4f99d63bd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_042653/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "500" +os.environ["TABDIFF_STEPS"] = "500" +os.environ["TABDIFF_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.0005" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0005" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "100" +os.environ["TABDIFF_TIMESTEPS"] = "100" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..f839c0318ac44e9e8e8b4a5888706d6b3d6b6723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb61666faaa53ca15d75e79a67d652f7bc157287f9e2f81dc2fe0a6d4545733f +size 127 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/gen_20260521_043250.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/gen_20260521_043250.log new file mode 100644 index 0000000000000000000000000000000000000000..42e6a0c952d90f30330cdffc2d3e850a634c3405 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/gen_20260521_043250.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7328f91a499318ec32538dfc5b815655c711f01652e90ea2ce1a324b77d952a0 +size 31805 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4b03f47cfcfd40e5faaffb6338bb729659fc60e8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6513c6c6c13cfe0e5e9458e028c32cedd479221367f0c89de595200f5fb0a9fd +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..30f4030a256d2900a1527af1b8f06c506c5eb98e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8169aba67b008ecb5a9d57ddc71a6614c374bddb600688b8f8d3312ee9b6bf0 +size 2435 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..1d91cd7d5d78b0a9d0dd20a917a0cca8f3d0b18c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c138bcbf2ecb43926d30784f6fe3ae91c90409774c5746cddec9ca03511c22da +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..606af493a7631c13fcc169f9982fcd2e624d0c75 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbfacaf711ca98021de964fc3c70e3973322c3bb580d988aff32dd3f9733663a +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0b2e2a20d420449285c3a86836f78d9e96980e91 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d1b0807356d40bcff37fb290b161b9afc3570557b8d075c46d8cb9074ce4c7f +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabdiff-n11-15215-20260521_043250.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabdiff-n11-15215-20260521_043250.csv new file mode 100644 index 0000000000000000000000000000000000000000..3b6c1a4d0f822cb97c48ed6d254691436b8f0e44 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabdiff-n11-15215-20260521_043250.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee0bbb130e6c7691009c74083f0ffd6c4ddbb5aabbc5e73a4b36669b2e2cc4a1 +size 1555682 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..5a9f5a229ea2c8bf37ca43b92e992729c4e62ceb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:428e4ab60a32507c67fe26c0a48217265d36ba55541fe819ba2722500f83071b +size 536 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/train_20260521_042653.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/train_20260521_042653.log new file mode 100644 index 0000000000000000000000000000000000000000..d6f16aa468c4eced461eb59720bd58613c6254b9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_042653/train_20260521_042653.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da5b52afab5e61adfc9eb368f1b6c1fe90aae59b7ed45e60fb0d520e9e714ee6 +size 1621534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..8f4d00f53bdb7c8f246b4293c81673ee7859b275 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_043507/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_500.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_043507/tabdiff-n11-15215-20260521_044033.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..d1513b3d7a73789d5df9da9138e82041c43c2850 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_043507/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "500" +os.environ["TABDIFF_STEPS"] = "500" +os.environ["TABDIFF_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.0001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "50" +os.environ["TABDIFF_TIMESTEPS"] = "50" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..71c5a7641c5eed0b3f569a1ab06a6084e50a4920 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ceebe375e803879a9b2b8dae3245fb48ac1c170bc81328d54d283c62313b28f +size 126 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/gen_20260521_044033.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/gen_20260521_044033.log new file mode 100644 index 0000000000000000000000000000000000000000..f95129c2d488373c4d27493851af26b9e430cf26 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/gen_20260521_044033.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcc8f020cca81c65bf21dcfb437fdce42cc9ed42676b31176aa6bb2444ce6ed8 +size 20767 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..cbef245b33d7338509692ae9d4f9a5eb5e604e85 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b483fdf9e43c473c56873c6740eeb4dcda5c27e65aa4618a1ab851dfe8c81891 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..bbb779369e5b64d1d5a787b7f65c2fa2bff21386 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22a8c572c1a7a2f9d9635a1862658a4c73717f93e6202ab6f1a1b2b916947ca9 +size 2434 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..9d7d121a91e82f7147148979eb27a940fbdae19e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a88a8fb13abaf85651cceb93d2d9b46de2d8310ec5951d1a45df438cd2e7282a +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..b86b2fdff60e5f273c0e34344848cdc29bc563ec --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0e76c79918679d9477e826b7e9ddbbe09dd877245cb7b84479cb98d8913c8de +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..02e8821fb08211dd41c967420fb7513191ba8598 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f25b62f7c56a7cb37b71323f9a506dc3f602513e1239b100351e1bea64092f0c +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabdiff-n11-15215-20260521_044033.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabdiff-n11-15215-20260521_044033.csv new file mode 100644 index 0000000000000000000000000000000000000000..cbccc3ee86fda357f1449c764f3cbe49f3459d5c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabdiff-n11-15215-20260521_044033.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48592de4fd4a45fa022c4a10d07078921c3997cabd3334f42be73f7854ac6581 +size 1554763 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..45129e236cd9e45516834a36378ec3e296d80100 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac2b3a36df89b9ca9c1d68eab2ba1a265137c985f9b8064865c1d93910c3cc67 +size 535 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/train_20260521_043507.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/train_20260521_043507.log new file mode 100644 index 0000000000000000000000000000000000000000..f171fff63e05eb1bbfc87a2ef8212f09b385ece8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_043507/train_20260521_043507.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8c0c395a7599b5542956d8b9c2b2f2ba6437ed664772df5dc1cc09db70b9163 +size 1590108 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..0bd419c75f6c7d75246b61797f0c2420f8837d54 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_044212/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_800.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_044212/tabdiff-n11-15215-20260521_045741.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..6dba331386a5c43a9034fd6f969a4f9a0ae46e85 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_044212/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "800" +os.environ["TABDIFF_STEPS"] = "800" +os.environ["TABDIFF_BATCH_SIZE"] = "512" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "512" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.0001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "100" +os.environ["TABDIFF_TIMESTEPS"] = "100" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..079d4ed8ab31376c9ea90f33e81236b0737ed1ed --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:663d4c2c32c989a3abb1ad6e0d5360401b53e6c3c883fd3d8e2f2ad762cfd398 +size 126 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/gen_20260521_045741.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/gen_20260521_045741.log new file mode 100644 index 0000000000000000000000000000000000000000..f8f39c2e03d5eaa0c6c35daf39ac9dc5ea614aaf --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/gen_20260521_045741.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9cba888ba25426b3f70f4eee4d323f896d20c0e394b5b0d17369f684a27d510 +size 30156 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2f3fff2b5cf95d16f9a6e39da68ce50040d9ebf2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1a3f5f7af4bf32dc2a5a0e21646d310cc2109718d692cd65288d5e70c770088 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..e23caca872cbe799180d8291a6ad64cb4acfcf45 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96a1ad781a866a4203ff9c17bd93fda4f1dbff86367a4a08c517c33c433ce372 +size 2434 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..aa539c4dbc614f540dc0c1f03309da22750acac7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45cc6e96c5bcaa5fab6d80c3ab3b1a076a30cf7f2dd02ee73fa419f90fddaa0e +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..5974895738d032d13cf2b5a46ac397b8710b9ba0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18aec6ef481a8e555627f391315ef2c0d1793ba7b0a7a799d8d2d665f0286bce +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..5409e08ea026d74ea9ca8d8956e6453ec6d73d77 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39763b29a4ec80d2e685abeaeb2357bb48ea59e60aae2c8fe8d2de51e68351b2 +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabdiff-n11-15215-20260521_045741.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabdiff-n11-15215-20260521_045741.csv new file mode 100644 index 0000000000000000000000000000000000000000..894ca70144f796306723c671e31d48c644fc45cc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabdiff-n11-15215-20260521_045741.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8328cea894d4887bd32829840233fe7e699ba9cf87b60124166717c8c327c392 +size 1556172 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..b4835e21c2d4128c0dbd33d2ca97ae8da1f1fe9a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53efc15a5fcb5cb6ecb28511d7cdbe1d020132e6da9373a619167105c958dbab +size 535 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/train_20260521_044213.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/train_20260521_044213.log new file mode 100644 index 0000000000000000000000000000000000000000..a1aaab1f3e49ebe640357c31c5c207c55da4753a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_044212/train_20260521_044213.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5d0679c7b34f3557d4a08364ba6fd4249b577c5ee9e80636b275c659946fff3 +size 5008869 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..6e0600767a4f52ee1b9b8bd26ab85c818ff6a3a1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_045956/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_800.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_045956/tabdiff-n11-15215-20260521_051447.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..8710a4dc8220615cd21cdc04e3498dbb803697fc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_045956/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "800" +os.environ["TABDIFF_STEPS"] = "800" +os.environ["TABDIFF_BATCH_SIZE"] = "512" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "512" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "25" +os.environ["TABDIFF_TIMESTEPS"] = "25" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..7fd6743544507ac15039fefcb5b2908a9c38b911 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae15dada9f3509e8f8e7110da5f7a9387c39939d71a62efcbc30a416a2eddd6e +size 124 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/gen_20260521_051447.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/gen_20260521_051447.log new file mode 100644 index 0000000000000000000000000000000000000000..d20c9f2e0fc91700e99c8d8d3e1da5192647d762 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/gen_20260521_051447.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1379767022d29d0a1bc5f4b1189cb9111a8db78353f92da435a72b6008ca2e08 +size 15104 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..7d013660d8efa324419a1d0f9a68808fb1959701 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb0e47238b054cf328caf97bf8fe2b60ca70c69be84c05d1b35d715a6ef3ce76 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..90bb5e1017bd472ca412cdb21a769c4a8ba53d17 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fc843db5e986a20aee9031585a3e2f4ad286b9b11e112439c1620334d221aa7 +size 2432 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..a256c1d03361d1e369197c4ea65f74c012272718 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbf6f7701601be409d7853072e9d82910b11b9c5817acf34e20527f216d120d6 +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..68922b0bd65ce3393ce86ddc1884a1c22582ad5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb2bbb2f0e0efbaef8c17bda560607ff3c1829dc8133c877f1afb8ba8257c7c1 +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..7844f08158abd759bf77cd095e84717f66e8d1e1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7ff316916aef42ceec7a49794da39220d5c6f2a684c840d468925bee6ab7f91 +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabdiff-n11-15215-20260521_051447.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabdiff-n11-15215-20260521_051447.csv new file mode 100644 index 0000000000000000000000000000000000000000..716a855009b072574b55d7d801a7155fb891fb41 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabdiff-n11-15215-20260521_051447.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f5ceff7950ace6e5d7bd0ef37179cfd1ec105759ddcd4b708325965b8014dbe +size 1556143 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..7a21ba29a5b1cb4cb9889c1c1267ccedeb5d7f2a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9068680ce9798ae6ae6c4a61457456063ffc7cf1d4a251704680852c5aa3a1fb +size 533 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/train_20260521_045956.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/train_20260521_045956.log new file mode 100644 index 0000000000000000000000000000000000000000..8992eeaf27847400ec7e6883d70f50242ea806dd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_045956/train_20260521_045956.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26ef2e4e8bc5c8e79f7ca981cc3f0eab4e61d9fad3f385c403315452e9cea7c7 +size 5005398 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..a12c79c77c53408420a8a83db89612a7d791a5ad --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_051642/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_800.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_051642/tabdiff-n11-15215-20260521_053144.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..483fb2ee7e1db553e97a01e2337fa81653871fbc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_051642/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "800" +os.environ["TABDIFF_STEPS"] = "800" +os.environ["TABDIFF_BATCH_SIZE"] = "512" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "512" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.0005" +os.environ["TABDIFF_LEARNING_RATE"] = "0.0005" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "50" +os.environ["TABDIFF_TIMESTEPS"] = "50" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..179d78fe0c4d0fb6b293afbef37565d055d02ee8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d37a51ffaead2135bc00a1fcf46dba603bdd155db5a1b8825145bf6752cd7295 +size 125 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/gen_20260521_053144.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/gen_20260521_053144.log new file mode 100644 index 0000000000000000000000000000000000000000..b388024cda92832dc7062b81e71b8ff55d9b5b94 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/gen_20260521_053144.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72dc4ea01ab55238a60b5c117ce26c12312b79c88530c72afea059156766521 +size 20387 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d3f7f51c7524d7b6a5b92ad22aaf150e6518c499 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32d9c1a9eafb6c20f6b23d6b217a5ddfde71b19c8ae9e59dea50d011a252d99b +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..2f7e3f3490ea444fb9df78c4d6178274809f23ae --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:897059ca6095e411a9cd7d8dce6a5acbafa461b766807ba300b2c5dc45e720cf +size 2433 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..4a912a9aa8b2ab72cf922295bfa66642ed184fd2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b7f66df6e8dda0dbfd932327610396f9401f78a22d4472e3317547668906426 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..c1f20535e100dae2b96077932e56b60cca90c43b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd0d98b50a84988c33f6fa6e22b6e80992b440b8fab4e7b067a84e72b72ad321 +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d620c97474c6f57c0a8463aeeaeff47319200b1b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e3296beeba05b6b8688d4f5d2d1e14a40e1f9dab65700c6ff669c3560095cff +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabdiff-n11-15215-20260521_053144.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabdiff-n11-15215-20260521_053144.csv new file mode 100644 index 0000000000000000000000000000000000000000..0d38510c8f6e0e1596026c149417004f304dd039 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabdiff-n11-15215-20260521_053144.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61845644c05fdf5b2162f1dcc1ce55cbd87ab15a117e1c82de8f3be4cc95156c +size 1556916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..72d7d97746ab8464063944d8bdbafe5e3156bc18 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19e8c8193f6d065da2531a02d4a36c0f067daa5ae4c8c346b118fa2799a8a172 +size 534 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/train_20260521_051643.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/train_20260521_051643.log new file mode 100644 index 0000000000000000000000000000000000000000..2e36a8a16a9468e13936dc77531cc58d480b6d6c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_051642/train_20260521_051643.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5da1d3b70e43c003d4edbc75a11899ec8174fa5a06d49fa2642c63774d4b79ee +size 5034188 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_gen.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..dfecf1ed02ed0cee0ed8e03d58ace6a20ab58f4d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_gen.py @@ -0,0 +1,42 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_053348/_tabdiff_runtime" +if not os.path.exists(rt): + def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + shutil.copytree(td, rt, ignore=_ignore) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "test", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", + "--ckpt_path", r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/ckpt/pipeline_n11/adapter_learnable/model_800.pt", + "--num_samples_to_generate", str(int(15215)), +]) +# test() 写入 tabdiff/result////samples.csv +base = os.path.join(rt, "tabdiff", "result", name, r"adapter_learnable") +best = None +best_t = -1.0 +for root, _, files in os.walk(base): + if "samples.csv" in files: + p = os.path.join(root, "samples.csv") + t = os.path.getmtime(p) + if t > best_t: + best_t = t + best = p +if not best: + raise SystemExit("tabdiff: no samples.csv under " + base) +shutil.copy(best, r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_053348/tabdiff-n11-15215-20260521_054153.csv") diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/.gitignore b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a5ac347c2673508e8a3637b03a654e22fe11bbef --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/.gitignore @@ -0,0 +1,15 @@ +**/__pycache__/** +*.pyc + +data/* +!/data/Info/ + +wandb/ +eval/ +synthetic/ +impute/ +workspace/ +debug/ +tabdiff/result/ + +**/ckpt/* diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/LICENSE b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cda47b343e4c33bc2406f424cf7698c8ddf0c56b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/LICENSE @@ -0,0 +1,7 @@ +Copyright 2024 Minkai Xu + +Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/README.md b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7b05f40ea4922d2858c845f15ef0633a79a7eba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/README.md @@ -0,0 +1,201 @@ +# TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation + +

+ + MIT License + + + Openreview + + + Paper URL + +

+ +
+ Model Logo +

Figure 1: Visualing the generative process of TabDiff. A high-quality version of this video can be found at tabdiff_demo.mp4

+
+ +This repository provides the official implementation of TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025). + +## Latest Update +- [2025.04]：The categorical-heavy dataset **[Diabetes](https://archive.ics.uci.edu/dataset/296/diabetes+130-us+hospitals+for+years+1999-2008)** evaluated in the paper has now been released! +- [2025.02]：Our code is finally released! We have released part of the tested datasets. The rest will be released soon! + +## Introduction + +
+ Model Logo +

Figure 2: The high-level schema of TabDiff

+
+TabDiff is a unified diffusion framework designed to model all muti-modal distributions of tabular data in a single model. Its key innovations include: + +1) Framing the joint diffusion process in continuous time, +2) A feature-wised learnable diffusion process that offsets the heterogeneity across different feature distributions, +3) Classifier-free guidance conditional generation for missing column value imputation. + +The schema of TabDiff is presented in the figure above. For more details, please refer to [our paper](https://arxiv.org/abs/2410.20626). + + +## Environment Setup + +Create the main environment with [tabdiff.yaml](tabdiff.yaml). This environment will be used for all tasks except for the evaluation of additional data fidelity metrics (i.e., $\alpha$-precision and $\beta$-recall scores) + +``` +conda env create -f tabdiff.yaml +``` + +Create another environment with [synthcity.yaml](synthcity.yaml) to evaluate additional data fidelity metrics + +``` +conda env create -f synthcity.yaml +``` + +## Datasets Preparation + +### Using the datasets experimented in the paper + +Download raw datasets: + +``` +python download_dataset.py +``` + +Process datasets: + +``` +python process_dataset.py +``` + +### Using your own dataset + +First, create a directory for your dataset in [./data](./data): +``` +cd data +mkdir +``` + +Compile your raw tabular data in .csv format. **The first row should be the header** indicating the name of each column, and the remaining rows are records. After finishing these steps, place you data's csv file in the directory you just created and name it as .csv. + +Then, create .json in [./data/Info](./data/Info). Write this file with the metadata of your dataset, covering the following information: +``` +{ + "name": "", + "task_type": "[NAME_OF_TASK]", # binclass or regression + "header": "infer", + "column_names": null, + "num_col_idx": [LIST], # list of indices of numerical columns + "cat_col_idx": [LIST], # list of indices of categorical columns + "target_col_idx": [list], # list of indices of the target columns (for MLE) + "file_type": "csv", + "data_path": "data//.csv" + "test_path": null, +} +``` + +### Important Notes When Creating the Info File +- The MLE evaluation and the imputation task (see later sections for details) assume that one column of your data is the regression or classification target. To enable these tasks, you will need to specify `target_col_idx`. If you don't need to evalute MLE, you can comment out the following line: https://github.com/MinkaiXu/TabDiff/blob/0c4fc3bbfa19046d36c5dce64628df52d5c73d15/tabdiff/main.py#L152 +- The fields `target_col_idx`, `num_col_idx` and `cat_col_idx` must be multually exclusive—no column should appear in more than one of these lists. +- Set the task_type to "regression" if the target column is numerical, or "binclass" if it is categorical. + +Finally, run the following command to process your dataset: +``` +python process_dataset.py --dataname +``` + +## Training TabDiff + +To train an unconditional TabDiff model across the entire table, run + +``` +python main.py --dataname --mode train +``` + +Current Options of `````` are: adult, default, shoppers, magic, beijing, news + +Wanb logging is enabled by default. To disable it and log locally, add the ```--no_wandb``` flag. + +To disable the learnable noise schedules, add the ```--non_learnable_schedule```. Please note that in order for the code to test/sample from such model properly, you need to add this flag for all commands below. + +To specify your own experiment name, which will be used for logging and saving files, add ```--exp_name ```. This flag overwrites the default experiment name (learnable_schedule/non_learnable_schedule), so, similar to ```--non_learnable_schedule```, once added to training, you need to add it to all following commands as well. + +## Sampling and Evaluating TabDiff (Density, MLE, C2ST) + +To sample synthetic tables from trained TabDiff models and evaluate them, run +``` +python main.py --dataname --mode test --report --no_wandb +``` + +This will sample 20 synthetic tables randomly. Meanwhile, it will evaluate the density, mle, and c2st scores for each sample and report their average and standard deviation. The results will be printed out in the terminal, and the samples and detailed evaluation results will be placed in ./eval/report_runs///. + +## Evaluating on Additional Fidelity Metrics ($\alpha$-precision and $\beta$-recall scores) +To evaluate TabDiff on the additional fidelity metrics ($\alpha$-precision and $\beta$-recall scores), you need to first make sure that you have already generated some samples by the previous commands. Then, you need to switch to the `synthcity` environment (as the synthcity packet used to compute those metrics conflicts with the main environment), by running +``` +conda activate synthcity +``` +Then, evaluate the metrics by running +``` +python eval/eval_quality.py --dataname +``` + +Similarly, the results will be printed out in the terminal and added to ./eval/report_runs/// + +## Evaluating Data Privacy (DCR score) +To evalute the privacy metric DCR score, you first need to retrain all the models, as the metric requires an equal split between the training and testing data (our initial splits employ a 90/10 ratio). To retrain with an equal split, run the training command but append `_dcr` to `````` +``` +python main.py --dataname _dcr --mode train +``` + +Then, test the models on DCR with the same `_dcr` suffix +``` +python main.py --dataname _dcr --mode test --report --no_wandb +``` + + + +## Missing Value Imputation with Classifier-free Guidance (CFG) +Our current experiments only include imputing the target column. However, our implementation, located at ```sample_impute()``` in [unified_ctime_diffusion.py](./tabdiff/models/unified_ctime_diffusion.py), should support imputing multiple columns with different data types. + +### Training Guidance Model +In order to enable classifier-free guidance (CFG), you need to first train an unconditional guidance model on the target column by running the training command with the `--y_only` flag +``` +python main.py --dataname --mode train --y_only +``` + +### Sampling Imputed Tables +With the trained guidance model, you can then impute the missing target column by running the testing command with the `--impute` flag +``` +python main.py --dataname --mode test --impute --no_wandb +``` +This will, by default, randomly produce 50 imputed tables and save them to ./impute//. + +### Evaluating Imputation +You can then evaluate the imputation quality by running +``` +python eval_impute.py --dataname +``` + +## License + +This work is licensed undeer the MIT License. + +## Acknowledgement +This repo is built upon the previous work TabSyn's [[codebase]](https://github.com/amazon-science/tabsyn). Many thanks to Hengrui! + +## Citation +Please consider citing our work if you find it helpful in your research! +``` +@inproceedings{ +shi2025tabdiff, +title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, +author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, +booktitle={The Thirteenth International Conference on Learning Representations}, +year={2025}, +url={https://openreview.net/forum?id=swvURjrt8z} +} +``` +## Contact +If you encounter any problem, please file an issue on this GitHub repo. + +If you have any question regarding the paper, please contact Minkai at [minkai@stanford.edu](minkai@stanford.edu) or Juntong at [shisteve@usc.edu](shisteve@usc.edu). diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/download_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/download_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d86e43dd3aa876fc659059baad90401bd3d3113f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/download_dataset.py @@ -0,0 +1,51 @@ +import os +from urllib import request +import zipfile + +DATA_DIR = 'data' + + +NAME_URL_DICT_UCI = { + 'adult': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'news_nocat': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', + 'adult_dcr': 'https://archive.ics.uci.edu/static/public/2/adult.zip', + 'default_dcr': 'https://archive.ics.uci.edu/static/public/350/default+of+credit+card+clients.zip', + 'magic_dcr': 'https://archive.ics.uci.edu/static/public/159/magic+gamma+telescope.zip', + 'shoppers_dcr': 'https://archive.ics.uci.edu/static/public/468/online+shoppers+purchasing+intention+dataset.zip', + 'beijing_dcr': 'https://archive.ics.uci.edu/static/public/381/beijing+pm2+5+data.zip', + 'news_dcr': 'https://archive.ics.uci.edu/static/public/332/online+news+popularity.zip', + 'diabetes_dcr': 'https://archive.ics.uci.edu/static/public/296/diabetes+130-us+hospitals+for+years+1999-2008.zip', +} + +def unzip_file(zip_filepath, dest_path): + with zipfile.ZipFile(zip_filepath, 'r') as zip_ref: + zip_ref.extractall(dest_path) + + +def download_from_uci(name): + + print(f'Start processing dataset {name} from UCI.') + save_dir = f'{DATA_DIR}/{name}' + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + url = NAME_URL_DICT_UCI[name] + request.urlretrieve(url, f'{save_dir}/{name}.zip') + print(f'Finish downloading dataset from {url}, data has been saved to {save_dir}.') + + unzip_file(f'{save_dir}/{name}.zip', save_dir) + print(f'Finish unzipping {name}.') + + else: + print('Aready downloaded.') + +if __name__ == '__main__': + for name in NAME_URL_DICT_UCI.keys(): + download_from_uci(name) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/eval_impute.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/eval_impute.py new file mode 100644 index 0000000000000000000000000000000000000000..576a4340314af6c9eb9253d5f4e040179feda3e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/eval_impute.py @@ -0,0 +1,82 @@ +import numpy as np +import pandas as pd +from sklearn.preprocessing import OneHotEncoder +from sklearn.metrics import f1_score, roc_auc_score +from sklearn.metrics import root_mean_squared_error +import argparse +import json + + +parser = argparse.ArgumentParser(description='Missing Value Imputation') + +parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') +parser.add_argument('--exp_name', type=str, default=None) +parser.add_argument('--col', type=int, default=0, help='Numerical Column to Impute') +parser.add_argument('--non_learnable_schedule', action='store_true') + +args = parser.parse_args() + +dataname = args.dataname +exp_name = args.exp_name +if exp_name is None: + exp_name = "non_learnable_schedule" if args.non_learnable_schedule else "learnable_schedule" +col = args.col + +dataname = args.dataname + +data_dir = f'data/{dataname}' + +real_path = f'{data_dir}/test.csv' + +info_path = f'data/{dataname}/info.json' +with open(info_path, 'r') as f: + info = json.load(f) +task_type = info['task_type'] + + +encoder = OneHotEncoder() + +real_data = pd.read_csv(real_path) +target_col = real_data.columns[info['target_col_idx'][0]] + +if task_type == "binclass": + real_target = real_data[target_col].to_numpy().reshape(-1,1) + real_y = encoder.fit_transform(real_target).toarray() + + syn_y = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + target = syn_data[target_col].to_numpy().reshape(-1, 1) + syn_y.append(encoder.transform(target).toarray()) + + syn_y_prob = np.stack(syn_y).mean(0) + syn_y_oh = np.argmax(syn_y_prob, axis=1) + num_classes = np.max(syn_y_oh) + 1 + syn_y_oh = np.eye(num_classes)[syn_y_oh] + + + + + micro_f1 = f1_score(real_y.argmax(axis=1), syn_y_prob.argmax(axis=1), average='micro') + auc = roc_auc_score(real_y, syn_y_prob, average='micro') + auc_argmaxed = roc_auc_score(real_y, syn_y_oh, average='micro') + print("AUC: ", round(auc*100, 3)) +else: + y_test = real_data[target_col].to_numpy() + y_test = np.log(np.clip(y_test, 1, 20000)) + + syn_y_ = [] + error = [] + for i in range(50): + syn_path = f'impute/{dataname}/{exp_name}/{i}.csv' + syn_data = pd.read_csv(syn_path) + syn_y = syn_data[target_col].to_numpy() + syn_y = np.log(np.clip(syn_y, 1, 20000)) + syn_y_.append(syn_y) + + pred = np.stack(syn_y_).mean(0) + root_mean_squared = root_mean_squared_error(y_test, pred) # mean_squared_error with squared=False is deprecated + + print("RMSE:", round(root_mean_squared, 4)) + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/images/tabdiff_demo.gif b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/images/tabdiff_demo.gif new file mode 100644 index 0000000000000000000000000000000000000000..55c4a680d59620578982bd1a153bee9371a28d99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/images/tabdiff_demo.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3f9c4819c4ef4abf8ade862da25915b826212f3d3ef9e4774ddde19db88bf2 +size 3260845 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/images/tabdiff_demo.mp4 b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/images/tabdiff_demo.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..fafe19a6abe882d8471e9afb7e51aabb9b40f2c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/images/tabdiff_demo.mp4 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c8072ab9a7dcc73d7bc3ea32b0e961cae5fcd7d73b6ffc2ffb44e6c291962f +size 1108599 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/images/tabdiff_flowchart.jpg b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/images/tabdiff_flowchart.jpg new file mode 100644 index 0000000000000000000000000000000000000000..390f02d16c6271b81168d19954b56d354398309b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/images/tabdiff_flowchart.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e4115e21b32199dd6cc0a0664f924907fcf1f66e2063616c50025f597b7409 +size 382775 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/main.py new file mode 100644 index 0000000000000000000000000000000000000000..2da09d0287cf71849f54261d33a45b4be723573d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/main.py @@ -0,0 +1,46 @@ +import torch +from tabdiff.main import main as tabdiff_main +import argparse + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Training of TabDiff') + + # General configs + parser.add_argument('--dataname', type=str, default='adult', help='Name dataset, one of those in data/ dir') + parser.add_argument('--mode', type=str, default='train', help='train or test') + parser.add_argument('--method', type=str, default='tabdiff', help='Currently we only release our model TabDiff. Baselines will be released soon.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index') + parser.add_argument('--debug', action='store_true', help='Enable debug mode') + parser.add_argument('--no_wandb', action='store_true', help='disable wandb') + parser.add_argument('--exp_name', type=str, default=None, help='Experiment name, used to name log directories and the wandb run name') + parser.add_argument('--deterministic', action='store_true', help='Whether to make the entire process deterministic, i.e., fix global random seeds') + + # Configs for tabdiff + parser.add_argument('--y_only', action='store_true', help='Train guidance model that only model the target column') + parser.add_argument('--non_learnable_schedule', action='store_true', help='disable learnable noise schedule') + + # Configs for testing tabdiff + parser.add_argument('--num_samples_to_generate', type=int, default=None, help='Number of samples to be generated while testing') + parser.add_argument('--ckpt_path', type=str, default=None, help='Path to the model checkpoint to be tested') + parser.add_argument('--report', action='store_true', help="Report testing mode: this mode sequentially runs test runs and report the avg and std") + parser.add_argument('--num_runs', type=int, default=20, help="Number of runs to be averaged in the report testing mode") + + # Configs for imputation + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=50) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default="x_t") + parser.add_argument('--y_only_model_path', type=str, default=None, help="Path to the y_only model checkpoint that will be used as the unconditional guidance model") + parser.add_argument('--w_num', type=float, default=0.6) + parser.add_argument('--w_cat', type=float, default=0.6) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + tabdiff_main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/process_dataset.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/process_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d866c68416c3f627b0ddbcf94a5f337e0cd84201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/process_dataset.py @@ -0,0 +1,646 @@ +import numpy as np +import pandas as pd +import os +import sys +import json +import argparse + +from sklearn.preprocessing import OrdinalEncoder +from sklearn import model_selection + +TYPE_TRANSFORM ={ + 'float', np.float32, + 'str', str, + 'int', int +} + +INFO_PATH = 'data/Info' + +parser = argparse.ArgumentParser(description='process dataset') + +# General configs +parser.add_argument('--dataname', type=str, default=None, help='Name of dataset.') +args = parser.parse_args() + +def preprocess_beijing(): + with open(f'{INFO_PATH}/beijing.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_beijing_dcr(): + with open(f'{INFO_PATH}/beijing_dcr.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + + data_df = pd.read_csv(data_path) + columns = data_df.columns + + data_df = data_df[columns[1:]] + + + df_cleaned = data_df.dropna() + df_cleaned.to_csv(info['data_path'], index = False) + +def preprocess_news(remove_cat=False): + name = 'news' if not remove_cat else 'news_nocat' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_news_dcr(remove_cat=False): + name = 'news_dcr' if not remove_cat else 'news_nocat_dcr' + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['raw_data_path'] + data_df = pd.read_csv(data_path) + data_df = data_df.drop('url', axis=1) + + columns = np.array(data_df.columns.tolist()) + + cat_columns1 = columns[list(range(12,18))] + cat_columns2 = columns[list(range(30,38))] + + if not remove_cat: + cat_col1 = data_df[cat_columns1].astype(int).to_numpy().argmax(axis = 1) + cat_col2 = data_df[cat_columns2].astype(int).to_numpy().argmax(axis = 1) + + data_df = data_df.drop(cat_columns2, axis=1) + data_df = data_df.drop(cat_columns1, axis=1) + + if not remove_cat: + data_df['data_channel'] = cat_col1 + data_df['weekday'] = cat_col2 + + data_save_path = f'data/{name}/{name}.csv' + data_df.to_csv(f'{data_save_path}', index = False) + + columns = np.array(data_df.columns.tolist()) + num_columns = columns[list(range(45))] + cat_columns = ['data_channel', 'weekday'] if not remove_cat else [] + target_columns = columns[[45]] + + info['num_col_idx'] = list(range(45)) + info['cat_col_idx'] = [46, 47] if not remove_cat else [] + info['target_col_idx'] = [45] + info['data_path'] = data_save_path + + with open(f'{INFO_PATH}/{name}.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.2, 0.2 + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes.json', 'w') as file: + json.dump(info, file, indent=4) + +def preprocess_diabetes_dcr(): + """ + Preprocesses the diabetes dataset is aligned with the concurrent work + Continuous Diffusion for Mixed-Type Tabular Data (CDTD): + https://github.com/muellermarkus/cdtd + """ + with open(f'{INFO_PATH}/diabetes_dcr.json', 'r') as f: + info = json.load(f) + + info['num_col_idx'] = list(range(9)) + info['cat_col_idx'] = list(range(9, 36)) + info['target_col_idx'] = [36] + + data_path = info['raw_data_path'] + df = pd.read_csv(data_path, sep=',') + df = df[info['column_names']] + df = df.replace(r' ', np.nan) + df = df.replace(r'?', np.nan) + df = df.replace(r'', np.nan) + + num_features = [info['column_names'][idx] for idx in info['num_col_idx']] + cat_features = [info['column_names'][idx] for idx in info['cat_col_idx']] + target = info['column_names'][info['target_col_idx'][0]] + df[target] = np.where(df[target] == 'NO', 0, 1) + enc = OrdinalEncoder() + df['age'] = enc.fit_transform(df['age'].to_numpy().reshape(-1,1)) + + # remove rows with missings in targets + idx_target_nan = df[target].isna().to_numpy().nonzero()[0] + df.drop(labels = idx_target_nan, axis = 0, inplace = True) + + # for categorical features, replace missings with 'empty', which will be counted as a new category + df[cat_features] = df[cat_features].fillna('empty') + + # for continuous data, drop missing + df.dropna(inplace = True) + + # ensure correct types + X_cat = df[cat_features].to_numpy().astype('str') + X_cont = df[num_features].to_numpy().astype('float') + y = df[[target]].to_numpy() + + val_prop, test_prop = 0.0, 0.5 # 50-50 split for dcr eval + prop = val_prop / (1 - test_prop) + + stratify = None if info['task_type'] == 'regression' else y + X_cat_train, X_cat_test, X_cont_train, X_cont_test, y_train, y_test = \ + model_selection.train_test_split(X_cat, X_cont, y, test_size = test_prop, + stratify = stratify, random_state = 42) + if val_prop > 0: + stratify = None if info['task_type'] == 'regression' else y_train + X_cat_train, X_cat_val, X_cont_train, X_cont_val, y_train, y_val = \ + model_selection.train_test_split(X_cat_train, X_cont_train, y_train, + stratify = stratify, test_size = prop, + random_state = 42) + + train_df = pd.DataFrame(np.concatenate([X_cont_train, X_cat_train, y_train], axis = 1), columns = num_features + cat_features + [target]) + if val_prop > 0: + val_df = pd.DataFrame(np.concatenate([X_cont_val, X_cat_val, y_val], axis = 1), columns = num_features + cat_features + [target]) + else: + val_df = pd.DataFrame(columns = num_features + cat_features + [target]).astype(train_df.dtypes) + test_df = pd.DataFrame(np.concatenate([X_cont_test, X_cat_test, y_test], axis = 1), columns = num_features + cat_features + [target]) + + # Save the splited data + train_df.to_csv(info['data_path'], index = False) + val_df.to_csv(info['val_path'], index = False) + test_df.to_csv(info['test_path'], index = False) + # Save updated info + with open(f'{INFO_PATH}/diabetes_dcr.json', 'w') as file: + json.dump(info, file, indent=4) + + + +def get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names = None): + + if not column_names: + column_names = np.array(data_df.columns.tolist()) + + + idx_mapping = {} + + curr_num_idx = 0 + curr_cat_idx = len(num_col_idx) + curr_target_idx = curr_cat_idx + len(cat_col_idx) + + for idx in range(len(column_names)): + + if idx in num_col_idx: + idx_mapping[int(idx)] = curr_num_idx + curr_num_idx += 1 + elif idx in cat_col_idx: + idx_mapping[int(idx)] = curr_cat_idx + curr_cat_idx += 1 + else: + idx_mapping[int(idx)] = curr_target_idx + curr_target_idx += 1 + + + inverse_idx_mapping = {} + for k, v in idx_mapping.items(): + inverse_idx_mapping[int(v)] = k + + idx_name_mapping = {} + + for i in range(len(column_names)): + idx_name_mapping[int(i)] = column_names[i] + + return idx_mapping, inverse_idx_mapping, idx_name_mapping + + +def train_val_test_split(data_df, cat_columns, num_train = 0, num_test = 0): + total_num = data_df.shape[0] + idx = np.arange(total_num) + + + seed = 1234 + + while True: + np.random.seed(seed) + np.random.shuffle(idx) + + train_idx = idx[:num_train] + test_idx = idx[-num_test:] + + + train_df = data_df.loc[train_idx] + test_df = data_df.loc[test_idx] + + + + flag = 0 + for i in cat_columns: + if len(set(train_df[i])) != len(set(data_df[i])): + flag = 1 + break + + if flag == 0: + break + else: + seed += 1 + + return train_df, test_df, seed + + +def process_data(name): + + if name == 'news': + preprocess_news() + elif name == 'news_nocat': + preprocess_news(remove_cat=True) + elif name == 'news_dcr': + preprocess_news_dcr() + elif name == 'beijing': + preprocess_beijing() + elif name == 'beijing_dcr': + preprocess_beijing_dcr() + elif name == 'diabetes': + preprocess_diabetes() + elif name == 'diabetes_dcr': + preprocess_diabetes_dcr() + + with open(f'{INFO_PATH}/{name}.json', 'r') as f: + info = json.load(f) + + data_path = info['data_path'] + if info['file_type'] == 'csv': + data_df = pd.read_csv(data_path, header = info['header']) + + elif info['file_type'] == 'xls': + data_df = pd.read_excel(data_path, sheet_name='Data', header=1) + data_df = data_df.drop('ID', axis=1) + + num_data = data_df.shape[0] + + column_names = info['column_names'] if info['column_names'] else data_df.columns.tolist() + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + num_columns = [column_names[i] for i in num_col_idx] + cat_columns = [column_names[i] for i in cat_col_idx] + target_columns = [column_names[i] for i in target_col_idx] + + idx_mapping, inverse_idx_mapping, idx_name_mapping = get_column_name_mapping(data_df, num_col_idx, cat_col_idx, target_col_idx, column_names) + + has_val = bool(info['val_path']) + val_df = pd.DataFrame(columns=data_df.columns).astype(data_df.dtypes) # by default (val_path is not provided), set val_Df to be empty + if info['test_path']: + + # if testing data is given + test_path = info['test_path'] + + if "adult" in name: # BUG: currently data saved at adult's test_path cannot be directly loaded. Consider integrate the following code to a preprocesing function for adult + with open(test_path, 'r') as f: + lines = f.readlines()[1:] + test_save_path = f'data/{name}/test.data' + if not os.path.exists(test_save_path): + with open(test_save_path, 'a') as f1: + for line in lines: + save_line = line.strip('\n').strip('.') + f1.write(f'{save_line}\n') + + test_df = pd.read_csv(test_save_path, header = None) + else: + test_df = pd.read_csv(test_path, header = info['header']) + + if has_val: # currently you cannot have a val path without a test path + val_path = info['val_path'] + val_df = pd.read_csv(val_path, header = info['header']) + + train_df = data_df + + if "dcr" in name and "diabetes" not in name: # create 50/50 splits for dcr datasets; no need for this for diabetes dataset as it's done in preprocessing + complete_df = pd.concat([train_df, test_df, val_df], axis = 0, ignore_index=True) + num_data = complete_df.shape[0] + num_train = int(num_data*0.5) + num_test = num_data - num_train + complete_df.rename(columns = idx_name_mapping, inplace=True) + train_df, test_df, seed = train_val_test_split(complete_df, cat_columns, num_train, num_test) + + else: + # Train/ Test Split, 90% Training (50% for dcr eval exclusively), 10% Testing (Validation set will be selected from Training set) + if "dcr" in name: + num_train = int(num_data*0.5) + else: + num_train = int(num_data*0.9) + num_test = num_data - num_train + + train_df, test_df, seed = train_val_test_split(data_df, cat_columns, num_train, num_test) + + complete_df = pd.concat([train_df, test_df, val_df], axis = 0) + name_idx_mapping = {val: key for key, val in idx_name_mapping.items()} + int_columns = [] + int_col_idx = [] + int_col_idx_wrt_num = [] + for i, col_idx in enumerate(num_col_idx): + col = column_names[col_idx] + col_data = complete_df.iloc[:,col_idx] + is_int = (col_data%1 == 0).all() + if is_int: + int_columns.append(col) + int_col_idx.append(name_idx_mapping[col]) + int_col_idx_wrt_num.append(i) + info['int_col_idx'] = int_col_idx + info['int_columns'] = int_columns + info['int_col_idx_wrt_num'] = int_col_idx_wrt_num + + train_df.columns = range(len(train_df.columns)) + test_df.columns = range(len(test_df.columns)) + val_df.columns = range(len(val_df.columns)) + + print(name, train_df.shape, val_df.shape, test_df.shape, data_df.shape) + + col_info = {} + + for col_idx in num_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + + for col_idx in cat_col_idx: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + for col_idx in target_col_idx: + if info['task_type'] == 'regression': + col_info[col_idx] = {} + col_info['type'] = 'numerical' + col_info['max'] = float(train_df[col_idx].max()) + col_info['min'] = float(train_df[col_idx].min()) + else: + col_info[col_idx] = {} + col_info['type'] = 'categorical' + col_info['categorizes'] = list(set(train_df[col_idx])) + + info['column_info'] = col_info + + train_df.rename(columns = idx_name_mapping, inplace=True) + test_df.rename(columns = idx_name_mapping, inplace=True) + val_df.rename(columns = idx_name_mapping, inplace=True) + + for col in num_columns: + if (train_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (train_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + train_df.loc[train_df[col] == '?', col] = np.nan + for col in cat_columns: + train_df.loc[train_df[col] == '?', col] = 'nan' + for col in num_columns: + if (test_df[col] == ' ?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + if (test_df[col] == '?').sum() > 0: + print(col) + import pdb; pdb.set_trace() + test_df.loc[test_df[col] == '?', col] = np.nan + for col in cat_columns: + test_df.loc[test_df[col] == '?', col] = 'nan' + for col in num_columns: + val_df.loc[val_df[col] == '?', col] = np.nan + for col in cat_columns: + val_df.loc[val_df[col] == '?', col] = 'nan' + + if train_df.isna().any().any(): + print("Training data contains nan in the numerical cols") + import pdb; pdb.set_trace() + + + + X_num_train = train_df[num_columns].to_numpy().astype(np.float32) + X_cat_train = train_df[cat_columns].to_numpy() + y_train = train_df[target_columns].to_numpy() + + + X_num_test = test_df[num_columns].to_numpy().astype(np.float32) + X_cat_test = test_df[cat_columns].to_numpy() + y_test = test_df[target_columns].to_numpy() + + X_num_val = val_df[num_columns].to_numpy().astype(np.float32) + X_cat_val = val_df[cat_columns].to_numpy() + y_val = val_df[target_columns].to_numpy() + + save_dir = f'data/{name}' + np.save(f'{save_dir}/X_num_train.npy', X_num_train) + np.save(f'{save_dir}/X_cat_train.npy', X_cat_train) + np.save(f'{save_dir}/y_train.npy', y_train) + + np.save(f'{save_dir}/X_num_test.npy', X_num_test) + np.save(f'{save_dir}/X_cat_test.npy', X_cat_test) + np.save(f'{save_dir}/y_test.npy', y_test) + + if has_val: + np.save(f'{save_dir}/X_num_val.npy', X_num_val) + np.save(f'{save_dir}/X_cat_val.npy', X_cat_val) + np.save(f'{save_dir}/y_val.npy', y_val) + + train_df[num_columns] = train_df[num_columns].astype(np.float32) + test_df[num_columns] = test_df[num_columns].astype(np.float32) + val_df[num_columns] = val_df[num_columns].astype(np.float32) + + + train_df.to_csv(f'{save_dir}/train.csv', index = False) + test_df.to_csv(f'{save_dir}/test.csv', index = False) + if has_val: + val_df.to_csv(f'{save_dir}/val.csv', index = False) + + if not os.path.exists(f'synthetic/{name}'): + os.makedirs(f'synthetic/{name}') + + train_df.to_csv(f'synthetic/{name}/real.csv', index = False) + test_df.to_csv(f'synthetic/{name}/test.csv', index = False) + + if has_val: + val_df.to_csv(f'synthetic/{name}/val.csv', index = False) + + print('Numerical', X_num_train.shape) + print('Categorical', X_cat_train.shape) + + info['column_names'] = column_names + info['train_num'] = train_df.shape[0] + info['test_num'] = test_df.shape[0] + info['val_num'] = val_df.shape[0] + + info['idx_mapping'] = idx_mapping + info['inverse_idx_mapping'] = inverse_idx_mapping + info['idx_name_mapping'] = idx_name_mapping + + metadata = {'columns': {}} + task_type = info['task_type'] + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + for i in num_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + for i in cat_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + + if task_type == 'regression': + + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'numerical' + metadata['columns'][i]['computer_representation'] = 'Float' + + else: + for i in target_col_idx: + metadata['columns'][i] = {} + metadata['columns'][i]['sdtype'] = 'categorical' + + info['metadata'] = metadata + + with open(f'{save_dir}/info.json', 'w') as file: + json.dump(info, file, indent=4) + + print(f'Processing and Saving {name} Successfully!') + + print(name) + print('Total', info['train_num'] + info['test_num']) + print('Train', info['train_num']) + print('Val', info['val_num']) + print('Test', info['test_num']) + if info['task_type'] == 'regression': + num = len(info['num_col_idx'] + info['target_col_idx']) + cat = len(info['cat_col_idx']) + else: + cat = len(info['cat_col_idx'] + info['target_col_idx']) + num = len(info['num_col_idx']) + print('Num', num) + print('Int', len(info['int_col_idx'])) + print('Cat', cat) + + +if __name__ == "__main__": + + if args.dataname: + process_data(args.dataname) + else: + for name in [ + 'adult', 'default', 'shoppers', 'magic', 'beijing', 'news', 'news_nocat', 'diabetes', + 'adult_dcr', + 'default_dcr', + 'shoppers_dcr', + 'beijing_dcr', + 'news_dcr', + 'diabetes_dcr' + ]: + process_data(name) + + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/__init__.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..07d7a4057f36847f7accf4ca148d590d8b4fc80f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/__init__.py @@ -0,0 +1,11 @@ +import torch +from icecream import install + +torch.set_num_threads(1) +install() + +from . import env # noqa +from .data import * # noqa +from .env import * # noqa +from .metrics import * # noqa +from .util import * # noqa \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/data.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/data.py new file mode 100644 index 0000000000000000000000000000000000000000..65ddd52afdb5225bbec25fb287eb271525fc4723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/data.py @@ -0,0 +1,780 @@ +import hashlib +from collections import Counter +from copy import deepcopy +from dataclasses import astuple, dataclass, replace +from importlib.resources import path +from pathlib import Path +from typing import Any, Literal, Optional, Union, cast, Tuple, Dict, List + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.pipeline import make_pipeline +import sklearn.preprocessing +import torch +import os +from category_encoders import LeaveOneOutEncoder +from sklearn.impute import SimpleImputer +from sklearn.preprocessing import StandardScaler +from scipy.spatial.distance import cdist + +from . import env, util +from .metrics import calculate_metrics as calculate_metrics_ +from .util import TaskType, load_json + +ArrayDict = Dict[str, np.ndarray] +TensorDict = Dict[str, torch.Tensor] + + +CAT_MISSING_VALUE = 'nan' +CAT_RARE_VALUE = '__rare__' +Normalization = Literal['standard', 'quantile', 'minmax'] +NumNanPolicy = Literal['drop-rows', 'mean'] +CatNanPolicy = Literal['most_frequent'] +CatEncoding = Literal['one-hot', 'counter'] +YPolicy = Literal['default'] +DEQUANT_DIST = Literal['uniform', 'beta', 'round', 'none'] + + +class StandardScaler1d(StandardScaler): + def partial_fit(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().partial_fit(X[:, None], *args, **kwargs) + + def transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().transform(X[:, None], *args, **kwargs).squeeze(1) + + def inverse_transform(self, X, *args, **kwargs): + assert X.ndim == 1 + return super().inverse_transform(X[:, None], *args, **kwargs).squeeze(1) + + +def get_category_sizes(X: Union[torch.Tensor, np.ndarray]) -> List[int]: + XT = X.T.cpu().tolist() if isinstance(X, torch.Tensor) else X.T.tolist() + return [len(set(x)) for x in XT] + + +@dataclass(frozen=False) +class Dataset: + X_num: Optional[ArrayDict] + X_cat: Optional[ArrayDict] + y: ArrayDict + int_col_idx_wrt_num: list + y_info: Dict[str, Any] + task_type: TaskType + n_classes: Optional[int] + + @classmethod + def from_dir(cls, dir_: Union[Path, str]) -> 'Dataset': + dir_ = Path(dir_) + splits = [k for k in ['train', 'test'] if dir_.joinpath(f'y_{k}.npy').exists()] + + def load(item) -> ArrayDict: + return { + x: cast(np.ndarray, np.load(dir_ / f'{item}_{x}.npy', allow_pickle=True)) # type: ignore[code] + for x in splits + } + + if Path(dir_ / 'info.json').exists(): + info = util.load_json(dir_ / 'info.json') + else: + info = None + return Dataset( + load('X_num') if dir_.joinpath('X_num_train.npy').exists() else None, + load('X_cat') if dir_.joinpath('X_cat_train.npy').exists() else None, + load('y'), + {}, + TaskType(info['task_type']), + info.get('n_classes'), + ) + + @property + def is_binclass(self) -> bool: + return self.task_type == TaskType.BINCLASS + + @property + def is_multiclass(self) -> bool: + return self.task_type == TaskType.MULTICLASS + + @property + def is_regression(self) -> bool: + return self.task_type == TaskType.REGRESSION + + @property + def n_num_features(self) -> int: + return 0 if self.X_num is None else self.X_num['train'].shape[1] + + @property + def n_cat_features(self) -> int: + return 0 if self.X_cat is None else self.X_cat['train'].shape[1] + + @property + def n_features(self) -> int: + return self.n_num_features + self.n_cat_features + + def size(self, part: Optional[str]) -> int: + return sum(map(len, self.y.values())) if part is None else len(self.y[part]) + + @property + def nn_output_dim(self) -> int: + if self.is_multiclass: + assert self.n_classes is not None + return self.n_classes + else: + return 1 + + def get_category_sizes(self, part: str) -> List[int]: + return [] if self.X_cat is None else get_category_sizes(self.X_cat[part]) + + def calculate_metrics( + self, + predictions: Dict[str, np.ndarray], + prediction_type: Optional[str], + ) -> Dict[str, Any]: + metrics = { + x: calculate_metrics_( + self.y[x], predictions[x], self.task_type, prediction_type, self.y_info + ) + for x in predictions + } + if self.task_type == TaskType.REGRESSION: + score_key = 'rmse' + score_sign = -1 + else: + score_key = 'accuracy' + score_sign = 1 + for part_metrics in metrics.values(): + part_metrics['score'] = score_sign * part_metrics[score_key] + return metrics + +def change_val(dataset: Dataset, val_size: float = 0.2): + # should be done before transformations + + y = np.concatenate([dataset.y['train'], dataset.y['val']], axis=0) + + ixs = np.arange(y.shape[0]) + if dataset.is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + + dataset.y['train'] = y[train_ixs] + dataset.y['val'] = y[val_ixs] + + if dataset.X_num is not None: + X_num = np.concatenate([dataset.X_num['train'], dataset.X_num['val']], axis=0) + dataset.X_num['train'] = X_num[train_ixs] + dataset.X_num['val'] = X_num[val_ixs] + + if dataset.X_cat is not None: + X_cat = np.concatenate([dataset.X_cat['train'], dataset.X_cat['val']], axis=0) + dataset.X_cat['train'] = X_cat[train_ixs] + dataset.X_cat['val'] = X_cat[val_ixs] + + return dataset + +def num_process_nans(dataset: Dataset, policy: Optional[NumNanPolicy]) -> Dataset: + + assert dataset.X_num is not None + nan_masks = {k: np.isnan(v) for k, v in dataset.X_num.items()} + if not any(x.any() for x in nan_masks.values()): # type: ignore[code] + # assert policy is None + print('No NaNs in numerical features, skipping') + return dataset + + assert policy is not None + if policy == 'drop-rows': + valid_masks = {k: ~v.any(1) for k, v in nan_masks.items()} + assert valid_masks[ + 'test' + ].all(), 'Cannot drop test rows, since this will affect the final metrics.' + new_data = {} + for data_name in ['X_num', 'X_cat', 'y']: + data_dict = getattr(dataset, data_name) + if data_dict is not None: + new_data[data_name] = { + k: v[valid_masks[k]] for k, v in data_dict.items() + } + dataset = replace(dataset, **new_data) + elif policy == 'mean': + new_values = np.nanmean(dataset.X_num['train'], axis=0) + X_num = deepcopy(dataset.X_num) + for k, v in X_num.items(): + num_nan_indices = np.where(nan_masks[k]) + v[num_nan_indices] = np.take(new_values, num_nan_indices[1]) + dataset = replace(dataset, X_num=X_num) + else: + assert util.raise_unknown('policy', policy) + return dataset + + +# Inspired by: https://github.com/yandex-research/rtdl/blob/a4c93a32b334ef55d2a0559a4407c8306ffeeaee/lib/data.py#L20 +def normalize( + X: ArrayDict, normalization: Normalization, seed: Optional[int], return_normalizer : bool = False +) -> ArrayDict: + X_train = X['train'] + if normalization == 'standard': + normalizer = sklearn.preprocessing.StandardScaler() + elif normalization == 'minmax': + normalizer = sklearn.preprocessing.MinMaxScaler() + elif normalization == 'quantile': + normalizer = sklearn.preprocessing.QuantileTransformer( + output_distribution='normal', + n_quantiles=max(min(X['train'].shape[0] // 30, 1000), 10), + subsample=int(1e9), + random_state=seed, + ) + # noise = 1e-3 + # if noise > 0: + # assert seed is not None + # stds = np.std(X_train, axis=0, keepdims=True) + # noise_std = noise / np.maximum(stds, noise) # type: ignore[code] + # X_train = X_train + noise_std * np.random.default_rng(seed).standard_normal( + # X_train.shape + # ) + else: + util.raise_unknown('normalization', normalization) + + normalizer.fit(X_train) + if return_normalizer: + return {k: normalizer.transform(v) for k, v in X.items()}, normalizer + return {k: normalizer.transform(v) for k, v in X.items()} + +class dequantizer: + def __init__( + self, + dequant_dist: DEQUANT_DIST, + int_col_idx_wrt_num: list, + int_dequant_factor: float, + # return_dequantizer: bool = False + ): + self.dequant_dist = dequant_dist + self.int_col_idx_wrt_num = int_col_idx_wrt_num + self.int_dequant_factor = int_dequant_factor + def transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = X_int+ np.random.uniform(size=X_int.shape) * self.int_dequant_factor + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = X_int + np.random.beta(self.int_dequant_factor, self.int_dequant_factor, size=X_int.shape) - 0.5 + elif self.dequant_dist in ['round', 'none']: + pass + return X + def inverse_transform(self, X): + X_int = X[:, self.int_col_idx_wrt_num] + if self.dequant_dist == 'uniform': + X[:, self.int_col_idx_wrt_num] = np.floor(X_int) + elif self.dequant_dist == 'beta': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'round': + X[:, self.int_col_idx_wrt_num] = np.rint(X_int) + elif self.dequant_dist == 'none': + pass + return X + + + # if return_dequantizer: + # return {k: transform(v) for k, v in X.items()}, inverse_transform + # return {k: transform(v) for k, v in X.items()} + +def cat_process_nans(X: ArrayDict, policy: Optional[CatNanPolicy]) -> ArrayDict: + assert X is not None + nan_masks = {k: v == CAT_MISSING_VALUE for k, v in X.items()} + if any(x.any() for x in nan_masks.values()): # type: ignore[code] + if policy is None: + X_new = X + elif policy == 'most_frequent': + imputer = SimpleImputer(missing_values=CAT_MISSING_VALUE, strategy=policy) # type: ignore[code] + imputer.fit(X['train']) + X_new = {k: cast(np.ndarray, imputer.transform(v)) for k, v in X.items()} + else: + util.raise_unknown('categorical NaN policy', policy) + else: + assert policy is None + X_new = X + return X_new + + +def cat_drop_rare(X: ArrayDict, min_frequency: float) -> ArrayDict: + assert 0.0 < min_frequency < 1.0 + min_count = round(len(X['train']) * min_frequency) + X_new = {x: [] for x in X} + for column_idx in range(X['train'].shape[1]): + counter = Counter(X['train'][:, column_idx].tolist()) + popular_categories = {k for k, v in counter.items() if v >= min_count} + for part in X_new: + X_new[part].append( + [ + (x if x in popular_categories else CAT_RARE_VALUE) + for x in X[part][:, column_idx].tolist() + ] + ) + return {k: np.array(v).T for k, v in X_new.items()} + + +def cat_encode( + X: ArrayDict, + encoding: Optional[CatEncoding], + y_train: Optional[np.ndarray], + seed: Optional[int], + return_encoder : bool = False +) -> Tuple[ArrayDict, bool, Optional[Any]]: # (X, is_converted_to_numerical) + if encoding != 'counter': + y_train = None + + # Step 1. Map strings to 0-based ranges + + if encoding is None: + unknown_value = np.iinfo('int64').max - 3 + oe = sklearn.preprocessing.OrdinalEncoder( + handle_unknown='use_encoded_value', # type: ignore[code] + unknown_value=unknown_value, # type: ignore[code] + dtype='int64', # type: ignore[code] + ).fit(X['train']) + encoder = make_pipeline(oe) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + max_values = X['train'].max(axis=0) + for part in X.keys(): + if part == 'train': continue + for column_idx in range(X[part].shape[1]): + X[part][X[part][:, column_idx] == unknown_value, column_idx] = ( + max_values[column_idx] + 1 + ) + if return_encoder: + return (X, False, encoder) + return (X, False) + + # Step 2. Encode. + + elif encoding == 'one-hot': + ohe = sklearn.preprocessing.OneHotEncoder( + handle_unknown='ignore', sparse_output=False, dtype=np.float32 # type: ignore[code] + ) + encoder = make_pipeline(ohe) + + # encoder.steps.append(('ohe', ohe)) + encoder.fit(X['train']) + X = {k: encoder.transform(v) for k, v in X.items()} + + elif encoding == 'counter': + assert y_train is not None + assert seed is not None + loe = LeaveOneOutEncoder(sigma=0.1, random_state=seed, return_df=False) + encoder.steps.append(('loe', loe)) + encoder.fit(X['train'], y_train) + X = {k: encoder.transform(v).astype('float32') for k, v in X.items()} # type: ignore[code] + if not isinstance(X['train'], pd.DataFrame): + X = {k: v.values for k, v in X.items()} # type: ignore[code] + else: + util.raise_unknown('encoding', encoding) + + if return_encoder: + return X, True, encoder # type: ignore[code] + return (X, True) + + +def build_target( + y: ArrayDict, policy: Optional[YPolicy], task_type: TaskType +) -> Tuple[ArrayDict, Dict[str, Any]]: + info: Dict[str, Any] = {'policy': policy} + if policy is None: + pass + elif policy == 'default': + if task_type == TaskType.REGRESSION: + mean, std = float(y['train'].mean()), float(y['train'].std()) + y = {k: (v - mean) / std for k, v in y.items()} + info['mean'] = mean + info['std'] = std + else: + util.raise_unknown('policy', policy) + return y, info + + +@dataclass(frozen=True) +class Transformations: + seed: int = 0 + normalization: Optional[Normalization] = None + num_nan_policy: Optional[NumNanPolicy] = None + cat_nan_policy: Optional[CatNanPolicy] = None + cat_min_frequency: Optional[float] = None + cat_encoding: Optional[CatEncoding] = None + y_policy: Optional[YPolicy] = 'default' + dequant_dist: Optional[DEQUANT_DIST] = None + int_dequant_factor: Optional[float] = 0.0 + + +def transform_dataset( + dataset: Dataset, + transformations: Transformations, + cache_dir: Optional[Path], + return_transforms: bool = False +) -> Dataset: + # WARNING: the order of transformations matters. Moreover, the current + # implementation is not ideal in that sense. + if cache_dir is not None: + transformations_md5 = hashlib.md5( + str(transformations).encode('utf-8') + ).hexdigest() + transformations_str = '__'.join(map(str, astuple(transformations))) + cache_path = ( + cache_dir / f'cache__{transformations_str}__{transformations_md5}.pickle' + ) + if cache_path.exists(): + cache_transformations, value = util.load_pickle(cache_path) + if transformations == cache_transformations: + print( + f"Using cached features: {cache_dir.name + '/' + cache_path.name}" + ) + return value + else: + raise RuntimeError(f'Hash collision for {cache_path}') + else: + cache_path = None + + if dataset.X_num is not None: + dataset = num_process_nans(dataset, transformations.num_nan_policy) + + num_transform = None + int_transform = None + cat_transform = None + X_num = dataset.X_num + + int_col_idx_wrt_num = dataset.int_col_idx_wrt_num + if X_num is not None and int_col_idx_wrt_num and transformations.dequant_dist is not None: + int_transform = dequantizer( + transformations.dequant_dist, + int_col_idx_wrt_num, + transformations.int_dequant_factor, + ) + X_num = {k: int_transform.transform(v) for k, v in X_num.items()} + + if X_num is not None and transformations.normalization is not None: + has_num = all([x.shape[1]>0 for x in dataset.X_num.values()]) + if has_num: + X_num, num_transform = normalize( + X_num, + transformations.normalization, + transformations.seed, + return_normalizer=True + ) + num_transform = num_transform + + if dataset.X_cat is None: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + # assert transformations.cat_encoding is None + X_cat = None + else: + has_cat = all([x.shape[1]>0 for x in dataset.X_cat.values()]) + if not has_cat: + assert transformations.cat_nan_policy is None + assert transformations.cat_min_frequency is None + X_cat = dataset.X_cat + for split in X_cat.keys(): # a patch to make sure that the empty array is transformed into int dtype + X_cat[split] = X_cat[split].astype(np.int64) + else: + X_cat = cat_process_nans(dataset.X_cat, transformations.cat_nan_policy) + + if transformations.cat_min_frequency is not None: + X_cat = cat_drop_rare(X_cat, transformations.cat_min_frequency) + X_cat, is_num, cat_transform = cat_encode( + X_cat, + transformations.cat_encoding, + dataset.y['train'], + transformations.seed, + return_encoder=True + ) + + if is_num: + X_num = ( + X_cat + if X_num is None + else {x: np.hstack([X_num[x], X_cat[x]]) for x in X_num} + ) + X_cat = None + + + y, y_info = build_target(dataset.y, transformations.y_policy, dataset.task_type) + + dataset = replace(dataset, X_num=X_num, X_cat=X_cat, y=y, y_info=y_info) + dataset.num_transform = num_transform + dataset.int_transform = int_transform + dataset.cat_transform = cat_transform + + if cache_path is not None: + util.dump_pickle((transformations, dataset), cache_path) + # if return_transforms: + # return dataset, num_transform, cat_transform + return dataset + + +def build_dataset( + path: Union[str, Path], + transformations: Transformations, + cache: bool +) -> Dataset: + path = Path(path) + dataset = Dataset.from_dir(path) + return transform_dataset(dataset, transformations, path if cache else None) + + +def prepare_tensors( + dataset: Dataset, device: Union[str, torch.device] +) -> Tuple[Optional[TensorDict], Optional[TensorDict], TensorDict]: + X_num, X_cat, Y = ( + None if x is None else {k: torch.as_tensor(v) for k, v in x.items()} + for x in [dataset.X_num, dataset.X_cat, dataset.y] + ) + if device.type != 'cpu': + X_num, X_cat, Y = ( + None if x is None else {k: v.to(device) for k, v in x.items()} + for x in [X_num, X_cat, Y] + ) + assert X_num is not None + assert Y is not None + if not dataset.is_multiclass: + Y = {k: v.float() for k, v in Y.items()} + return X_num, X_cat, Y + +############### +## DataLoader## +############### + +class TabDataset(torch.utils.data.Dataset): + def __init__( + self, dataset : Dataset, split : Literal['train', 'val', 'test'] + ): + super().__init__() + + self.X_num = torch.from_numpy(dataset.X_num[split]) if dataset.X_num is not None else None + self.X_cat = torch.from_numpy(dataset.X_cat[split]) if dataset.X_cat is not None else None + self.y = torch.from_numpy(dataset.y[split]) + + assert self.y is not None + assert self.X_num is not None or self.X_cat is not None + + def __len__(self): + return len(self.y) + + def __getitem__(self, idx): + out_dict = { + 'y': self.y[idx].long() if self.y is not None else None, + } + + x = np.empty((0,)) + if self.X_num is not None: + x = self.X_num[idx] + if self.X_cat is not None: + x = torch.cat([x, self.X_cat[idx]], dim=0) + return x.float(), out_dict + +def prepare_dataloader( + dataset : Dataset, + split : str, + batch_size: int, +): + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader( + torch_dataset, + batch_size=batch_size, + shuffle=(split == 'train'), + num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + while True: + yield from loader + +def prepare_torch_dataloader( + dataset : Dataset, + split : str, + shuffle : bool, + batch_size: int, +) -> torch.utils.data.DataLoader: + + torch_dataset = TabDataset(dataset, split) + loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0'))) + + return loader + +def dataset_from_csv(paths : Dict[str, str], cat_features, target, T): + assert 'train' in paths + y = {} + X_num = {} + X_cat = {} if len(cat_features) else None + for split in paths.keys(): + df = pd.read_csv(paths[split]) + y[split] = df[target].to_numpy().astype(float) + if X_cat is not None: + X_cat[split] = df[cat_features].to_numpy().astype(str) + X_num[split] = df.drop(cat_features + [target], axis=1).to_numpy().astype(float) + + dataset = Dataset(X_num, X_cat, y, {}, None, len(np.unique(y['train']))) + return transform_dataset(dataset, T, None) + +class FastTensorDataLoader: + """ + A DataLoader-like object for a set of tensors that can be much faster than + TensorDataset + DataLoader because dataloader grabs individual indices of + the dataset and calls cat (slow). + Source: https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6 + """ + def __init__(self, *tensors, batch_size=32, shuffle=False): + """ + Initialize a FastTensorDataLoader. + :param *tensors: tensors to store. Must have the same length @ dim 0. + :param batch_size: batch size to load. + :param shuffle: if True, shuffle the data *in-place* whenever an + iterator is created out of this object. + :returns: A FastTensorDataLoader. + """ + assert all(t.shape[0] == tensors[0].shape[0] for t in tensors) + self.tensors = tensors + + self.dataset_len = self.tensors[0].shape[0] + self.batch_size = batch_size + self.shuffle = shuffle + + # Calculate # batches + n_batches, remainder = divmod(self.dataset_len, self.batch_size) + if remainder > 0: + n_batches += 1 + self.n_batches = n_batches + def __iter__(self): + if self.shuffle: + r = torch.randperm(self.dataset_len) + self.tensors = [t[r] for t in self.tensors] + self.i = 0 + return self + + def __next__(self): + if self.i >= self.dataset_len: + raise StopIteration + batch = tuple(t[self.i:self.i+self.batch_size] for t in self.tensors) + self.i += self.batch_size + return batch + + def __len__(self): + return self.n_batches + +def prepare_fast_dataloader( + D : Dataset, + split : str, + batch_size: int +): + + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + dataloader = FastTensorDataLoader(X, batch_size=batch_size, shuffle=(split=='train')) + while True: + yield from dataloader + +def prepare_fast_torch_dataloader( + D : Dataset, + split : str, + batch_size: int +): + if D.X_cat is not None: + X = torch.from_numpy(np.concatenate([D.X_num[split], D.X_cat[split]], axis=1)).float() + else: + X = torch.from_numpy(D.X_num[split]).float() + y = torch.from_numpy(D.y[split]) + dataloader = FastTensorDataLoader(X, y, batch_size=batch_size, shuffle=(split=='train')) + return dataloader + +def round_columns(X_real, X_synth, columns): + for col in columns: + uniq = np.unique(X_real[:,col]) + dist = cdist(X_synth[:, col][:, np.newaxis].astype(float), uniq[:, np.newaxis].astype(float)) + X_synth[:, col] = uniq[dist.argmin(axis=1)] + return X_synth + +def concat_features(D : Dataset): + if D.X_num is None: + assert D.X_cat is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_cat.items()} + elif D.X_cat is None: + assert D.X_num is not None + X = {k: pd.DataFrame(v, columns=range(D.n_features)) for k, v in D.X_num.items()} + else: + X = { + part: pd.concat( + [ + pd.DataFrame(D.X_num[part], columns=range(D.n_num_features)), + pd.DataFrame( + D.X_cat[part], + columns=range(D.n_num_features, D.n_features), + ), + ], + axis=1, + ) + for part in D.y.keys() + } + + return X + +def concat_to_pd(X_num, X_cat, y): + if X_num is None: + return pd.concat([ + pd.DataFrame(X_cat, columns=list(range(X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + if X_cat is not None: + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(X_cat, columns=list(range(X_num.shape[1], X_num.shape[1] + X_cat.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + return pd.concat([ + pd.DataFrame(X_num, columns=list(range(X_num.shape[1]))), + pd.DataFrame(y, columns=['y']) + ], axis=1) + +def read_pure_data(path, split='train'): + y = np.load(os.path.join(path, f'y_{split}.npy'), allow_pickle=True) + X_num = None + X_cat = None + if os.path.exists(os.path.join(path, f'X_num_{split}.npy')): + X_num = np.load(os.path.join(path, f'X_num_{split}.npy'), allow_pickle=True) + if os.path.exists(os.path.join(path, f'X_cat_{split}.npy')): + X_cat = np.load(os.path.join(path, f'X_cat_{split}.npy'), allow_pickle=True) + + return X_num, X_cat, y + +def read_changed_val(path, val_size=0.2): + path = Path(path) + X_num_train, X_cat_train, y_train = read_pure_data(path, 'train') + X_num_val, X_cat_val, y_val = read_pure_data(path, 'val') + is_regression = load_json(path / 'info.json')['task_type'] == 'regression' + + y = np.concatenate([y_train, y_val], axis=0) + + ixs = np.arange(y.shape[0]) + if is_regression: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777) + else: + train_ixs, val_ixs = train_test_split(ixs, test_size=val_size, random_state=777, stratify=y) + y_train = y[train_ixs] + y_val = y[val_ixs] + + if X_num_train is not None: + X_num = np.concatenate([X_num_train, X_num_val], axis=0) + X_num_train = X_num[train_ixs] + X_num_val = X_num[val_ixs] + + if X_cat_train is not None: + X_cat = np.concatenate([X_cat_train, X_cat_val], axis=0) + X_cat_train = X_cat[train_ixs] + X_cat_val = X_cat[val_ixs] + + return X_num_train, X_cat_train, y_train, X_num_val, X_cat_val, y_val + +############# + +def load_dataset_info(dataset_dir_name: str) -> Dict[str, Any]: + path = Path("data/" + dataset_dir_name) + info = util.load_json(path / 'info.json') + info['size'] = info['train_size'] + info['val_size'] + info['test_size'] + info['n_features'] = info['n_num_features'] + info['n_cat_features'] + info['path'] = path + return info \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/env.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/env.py new file mode 100644 index 0000000000000000000000000000000000000000..0b9dd619f31450414a9b31585aa58b9121834fbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/env.py @@ -0,0 +1,39 @@ +""" +Have not used in TabDDPM project. +""" + +import datetime +import os +import shutil +import typing as ty +from pathlib import Path + +PROJ = Path('tab-ddpm/').absolute().resolve() +EXP = PROJ / 'exp' +DATA = PROJ / 'data' + + +def get_path(path: ty.Union[str, Path]) -> Path: + if isinstance(path, str): + path = Path(path) + if not path.is_absolute(): + path = PROJ / path + return path.resolve() + + +def get_relative_path(path: ty.Union[str, Path]) -> Path: + return get_path(path).relative_to(PROJ) + + +def duplicate_path( + src: ty.Union[str, Path], alternative_project_dir: ty.Union[str, Path] +) -> None: + src = get_path(src) + alternative_project_dir = get_path(alternative_project_dir) + dst = alternative_project_dir / src.relative_to(PROJ) + dst.parent.mkdir(parents=True, exist_ok=True) + if dst.exists(): + dst = dst.with_name( + dst.name + '_' + datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + ) + (shutil.copytree if src.is_dir() else shutil.copyfile)(src, dst) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..188a9be3f8e9d59b07688aa189a876660dbef7cd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/metrics.py @@ -0,0 +1,157 @@ +import enum +from typing import Any, Optional, Tuple, Dict, Union, cast +from functools import partial + +import numpy as np +import scipy.special +import sklearn.metrics as skm + +from . import util +from .util import TaskType + + +class PredictionType(enum.Enum): + LOGITS = 'logits' + PROBS = 'probs' + +class MetricsReport: + def __init__(self, report: dict, task_type: TaskType): + self._res = {k: {} for k in report.keys()} + if task_type in (TaskType.BINCLASS, TaskType.MULTICLASS): + self._metrics_names = ["acc", "f1"] + for k in report.keys(): + self._res[k]["acc"] = report[k]["accuracy"] + self._res[k]["f1"] = report[k]["macro avg"]["f1-score"] + if task_type == TaskType.BINCLASS: + self._res[k]["roc_auc"] = report[k]["roc_auc"] + self._metrics_names.append("roc_auc") + + elif task_type == TaskType.REGRESSION: + self._metrics_names = ["r2", "rmse"] + for k in report.keys(): + self._res[k]["r2"] = report[k]["r2"] + self._res[k]["rmse"] = report[k]["rmse"] + else: + raise "Unknown TaskType!" + + def get_splits_names(self) -> list[str]: + return self._res.keys() + + def get_metrics_names(self) -> list[str]: + return self._metrics_names + + def get_metric(self, split: str, metric: str) -> float: + return self._res[split][metric] + + def get_val_score(self) -> float: + return self._res["val"]["r2"] if "r2" in self._res["val"] else self._res["val"]["f1"] + + def get_test_score(self) -> float: + return self._res["test"]["r2"] if "r2" in self._res["test"] else self._res["test"]["f1"] + + def print_metrics(self) -> None: + res = { + "val": {k: np.around(self._res["val"][k], 4) for k in self._res["val"]}, + "test": {k: np.around(self._res["test"][k], 4) for k in self._res["test"]} + } + + print("*"*100) + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + + return res + +class SeedsMetricsReport: + def __init__(self): + self._reports = [] + + def add_report(self, report: MetricsReport) -> None: + self._reports.append(report) + + def get_mean_std(self) -> dict: + res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + res[split][metric] = [x.get_metric(split, metric) for x in self._reports] + + agg_res = {k: {} for k in ["train", "val", "test"]} + for split in self._reports[0].get_splits_names(): + for metric in self._reports[0].get_metrics_names(): + for k, f in [("count", len), ("mean", np.mean), ("std", np.std)]: + agg_res[split][f"{metric}-{k}"] = f(res[split][metric]) + self._res = res + self._agg_res = agg_res + + return agg_res + + def print_result(self) -> dict: + res = {split: {k: float(np.around(self._agg_res[split][k], 4)) for k in self._agg_res[split]} for split in ["val", "test"]} + print("="*100) + print("EVAL RESULTS:") + print("[val]") + print(res["val"]) + print("[test]") + print(res["test"]) + print("="*100) + return res + +def calculate_rmse( + y_true: np.ndarray, y_pred: np.ndarray, std = None) -> float: + rmse = skm.mean_squared_error(y_true, y_pred) ** 0.5 + if std is not None: + rmse *= std + return rmse + + +def _get_labels_and_probs( + y_pred: np.ndarray, task_type: TaskType, prediction_type: Optional[PredictionType] +) -> Tuple[np.ndarray, Optional[np.ndarray]]: + assert task_type in (TaskType.BINCLASS, TaskType.MULTICLASS) + + if prediction_type is None: + return y_pred, None + + if prediction_type == PredictionType.LOGITS: + probs = ( + scipy.special.expit(y_pred) + if task_type == TaskType.BINCLASS + else scipy.special.softmax(y_pred, axis=1) + ) + elif prediction_type == PredictionType.PROBS: + probs = y_pred + else: + util.raise_unknown('prediction_type', prediction_type) + + assert probs is not None + labels = np.round(probs) if task_type == TaskType.BINCLASS else probs.argmax(axis=1) + return labels.astype('int64'), probs + + +def calculate_metrics( + y_true: np.ndarray, + y_pred: np.ndarray, + task_type: Union[str, TaskType], + prediction_type: Optional[Union[str, PredictionType]], + y_info: Dict[str, Any], +) -> Dict[str, Any]: + # Example: calculate_metrics(y_true, y_pred, 'binclass', 'logits', {}) + task_type = TaskType(task_type) + if prediction_type is not None: + prediction_type = PredictionType(prediction_type) + + if task_type == TaskType.REGRESSION: + assert prediction_type is None + assert 'std' in y_info + rmse = calculate_rmse(y_true, y_pred, y_info['std']) + r2 = skm.r2_score(y_true, y_pred) + result = {'rmse': rmse, 'r2': r2} + else: + labels, probs = _get_labels_and_probs(y_pred, task_type, prediction_type) + result = cast( + Dict[str, Any], skm.classification_report(y_true, labels, output_dict=True) + ) + if task_type == TaskType.BINCLASS: + result['roc_auc'] = skm.roc_auc_score(y_true, probs) + return result \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/util.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e105c961803b3bf0f8282e673eed8e84eaaa3891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/src/util.py @@ -0,0 +1,347 @@ +import argparse +import atexit +import enum +import json +import os +import pickle +import shutil +import sys +import time +import uuid +from copy import deepcopy +from dataclasses import asdict, fields, is_dataclass +from pathlib import Path +from pprint import pprint +from typing import Any, Callable, List, Dict, Type, Optional, Tuple, TypeVar, Union, cast, get_args, get_origin + +import __main__ +import numpy as np +import tomli +import tomli_w +import torch +import typing as ty + +from . import env + +RawConfig = Dict[str, Any] +Report = Dict[str, Any] +T = TypeVar('T') + + +class Part(enum.Enum): + TRAIN = 'train' + VAL = 'val' + TEST = 'test' + + def __str__(self) -> str: + return self.value + + +class TaskType(enum.Enum): + BINCLASS = 'binclass' + MULTICLASS = 'multiclass' + REGRESSION = 'regression' + + def __str__(self) -> str: + return self.value + + + +def update_training_log(training_log, data, metrics): + def _update(log_part, data_part): + for k, v in data_part.items(): + if isinstance(v, dict): + _update(log_part.setdefault(k, {}), v) + elif isinstance(v, list): + log_part.setdefault(k, []).extend(v) + else: + log_part.setdefault(k, []).append(v) + + _update(training_log, data) + transposed_metrics = {} + for part, part_metrics in metrics.items(): + for metric_name, value in part_metrics.items(): + transposed_metrics.setdefault(metric_name, {})[part] = value + _update(training_log, transposed_metrics) + + +def raise_unknown(unknown_what: str, unknown_value: Any): + raise ValueError(f'Unknown {unknown_what}: {unknown_value}') + + +def _replace(data, condition, value): + def do(x): + if isinstance(x, dict): + return {k: do(v) for k, v in x.items()} + elif isinstance(x, list): + return [do(y) for y in x] + else: + return value if condition(x) else x + + return do(data) + + +_CONFIG_NONE = '__none__' + + +def unpack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x == _CONFIG_NONE, None)) + return config + + +def pack_config(config: RawConfig) -> RawConfig: + config = cast(RawConfig, _replace(config, lambda x: x is None, _CONFIG_NONE)) + return config + + +def load_config(path: Union[Path, str]) -> Any: + with open(path, 'rb') as f: + return unpack_config(tomli.load(f)) + + +def dump_config(config: Any, path: Union[Path, str]) -> None: + with open(path, 'wb') as f: + tomli_w.dump(pack_config(config), f) + # check that there are no bugs in all these "pack/unpack" things + assert config == load_config(path) + + +def load_json(path: Union[Path, str], **kwargs) -> Any: + return json.loads(Path(path).read_text(), **kwargs) + + +def dump_json(x: Any, path: Union[Path, str], **kwargs) -> None: + kwargs.setdefault('indent', 4) + Path(path).write_text(json.dumps(x, **kwargs) + '\n') + + +def load_pickle(path: Union[Path, str], **kwargs) -> Any: + return pickle.loads(Path(path).read_bytes(), **kwargs) + + +def dump_pickle(x: Any, path: Union[Path, str], **kwargs) -> None: + Path(path).write_bytes(pickle.dumps(x, **kwargs)) + + +def load(path: Union[Path, str], **kwargs) -> Any: + return globals()[f'load_{Path(path).suffix[1:]}'](Path(path), **kwargs) + + +def dump(x: Any, path: Union[Path, str], **kwargs) -> Any: + return globals()[f'dump_{Path(path).suffix[1:]}'](x, Path(path), **kwargs) + + +def _get_output_item_path( + path: Union[str, Path], filename: str, must_exist: bool +) -> Path: + path = env.get_path(path) + if path.suffix == '.toml': + path = path.with_suffix('') + if path.is_dir(): + path = path / filename + else: + assert path.name == filename + assert path.parent.exists() + if must_exist: + assert path.exists() + return path + + +def load_report(path: Path) -> Report: + return load_json(_get_output_item_path(path, 'report.json', True)) + + +def dump_report(report: dict, path: Path) -> None: + dump_json(report, _get_output_item_path(path, 'report.json', False)) + + +def load_predictions(path: Path) -> Dict[str, np.ndarray]: + with np.load(_get_output_item_path(path, 'predictions.npz', True)) as predictions: + return {x: predictions[x] for x in predictions} + + +def dump_predictions(predictions: Dict[str, np.ndarray], path: Path) -> None: + np.savez(_get_output_item_path(path, 'predictions.npz', False), **predictions) + + +def dump_metrics(metrics: Dict[str, Any], path: Path) -> None: + dump_json(metrics, _get_output_item_path(path, 'metrics.json', False)) + + +def load_checkpoint(path: Path, *args, **kwargs) -> Dict[str, np.ndarray]: + return torch.load( + _get_output_item_path(path, 'checkpoint.pt', True), *args, **kwargs + ) + + +def get_device() -> torch.device: + if torch.cuda.is_available(): + assert os.environ.get('CUDA_VISIBLE_DEVICES') is not None + return torch.device('cuda:0') + else: + return torch.device('cpu') + + +def _print_sep(c, size=100): + print(c * size) + + +_LAST_SNAPSHOT_TIME = None + + +def backup_output(output_dir: Path) -> None: + backup_dir = os.environ.get('TMP_OUTPUT_PATH') + snapshot_dir = os.environ.get('SNAPSHOT_PATH') + if backup_dir is None: + assert snapshot_dir is None + return + assert snapshot_dir is not None + + try: + relative_output_dir = output_dir.relative_to(env.PROJ) + except ValueError: + return + + for dir_ in [backup_dir, snapshot_dir]: + new_output_dir = dir_ / relative_output_dir + prev_backup_output_dir = new_output_dir.with_name(new_output_dir.name + '_prev') + new_output_dir.parent.mkdir(exist_ok=True, parents=True) + if new_output_dir.exists(): + new_output_dir.rename(prev_backup_output_dir) + shutil.copytree(output_dir, new_output_dir) + # the case for evaluate.py which automatically creates configs + if output_dir.with_suffix('.toml').exists(): + shutil.copyfile( + output_dir.with_suffix('.toml'), new_output_dir.with_suffix('.toml') + ) + if prev_backup_output_dir.exists(): + shutil.rmtree(prev_backup_output_dir) + + global _LAST_SNAPSHOT_TIME + if _LAST_SNAPSHOT_TIME is None or time.time() - _LAST_SNAPSHOT_TIME > 10 * 60: + import nirvana_dl.snapshot # type: ignore[code] + + nirvana_dl.snapshot.dump_snapshot() + _LAST_SNAPSHOT_TIME = time.time() + print('The snapshot was saved!') + + +def _get_scores(metrics: Dict[str, Dict[str, Any]]) -> Optional[Dict[str, float]]: + return ( + {k: v['score'] for k, v in metrics.items()} + if 'score' in next(iter(metrics.values())) + else None + ) + + +def format_scores(metrics: Dict[str, Dict[str, Any]]) -> str: + return ' '.join( + f"[{x}] {metrics[x]['score']:.3f}" + for x in ['test', 'val', 'train'] + if x in metrics + ) + + +def finish(output_dir: Path, report: dict) -> None: + print() + _print_sep('=') + + metrics = report.get('metrics') + if metrics is not None: + scores = _get_scores(metrics) + if scores is not None: + dump_json(scores, output_dir / 'scores.json') + print(format_scores(metrics)) + _print_sep('-') + + dump_report(report, output_dir) + json_output_path = os.environ.get('JSON_OUTPUT_FILE') + if json_output_path: + try: + key = str(output_dir.relative_to(env.PROJ)) + except ValueError: + pass + else: + json_output_path = Path(json_output_path) + try: + json_data = json.loads(json_output_path.read_text()) + except (FileNotFoundError, json.decoder.JSONDecodeError): + json_data = {} + json_data[key] = load_json(output_dir / 'report.json') + json_output_path.write_text(json.dumps(json_data, indent=4)) + shutil.copyfile( + json_output_path, + os.path.join(os.environ['SNAPSHOT_PATH'], 'json_output.json'), + ) + + output_dir.joinpath('DONE').touch() + backup_output(output_dir) + print(f'Done! | {report.get("time")} | {output_dir}') + _print_sep('=') + print() + + +def from_dict(datacls: Type[T], data: dict) -> T: + assert is_dataclass(datacls) + data = deepcopy(data) + for field in fields(datacls): + if field.name not in data: + continue + if is_dataclass(field.type): + data[field.name] = from_dict(field.type, data[field.name]) + elif ( + get_origin(field.type) is Union + and len(get_args(field.type)) == 2 + and get_args(field.type)[1] is type(None) + and is_dataclass(get_args(field.type)[0]) + ): + if data[field.name] is not None: + data[field.name] = from_dict(get_args(field.type)[0], data[field.name]) + return datacls(**data) + + +def replace_factor_with_value( + config: RawConfig, + key: str, + reference_value: int, + bounds: Tuple[float, float], +) -> None: + factor_key = key + '_factor' + if factor_key not in config: + assert key in config + else: + assert key not in config + factor = config.pop(factor_key) + assert bounds[0] <= factor <= bounds[1] + config[key] = int(factor * reference_value) + + +def get_temporary_copy(path: Union[str, Path]) -> Path: + path = env.get_path(path) + assert not path.is_dir() and not path.is_symlink() + tmp_path = path.with_name( + path.stem + '___' + str(uuid.uuid4()).replace('-', '') + path.suffix + ) + shutil.copyfile(path, tmp_path) + atexit.register(lambda: tmp_path.unlink()) + return tmp_path + + +def get_python(): + python = Path('python3.9') + return str(python) if python.exists() else 'python' + +def get_catboost_config(real_data_path, is_cv=False): + ds_name = Path(real_data_path).name + C = load_json(f'tuned_models/catboost/{ds_name}_cv.json') + return C + +def get_categories(X_train_cat): + return ( + None + if X_train_cat is None + else [ + len(set(X_train_cat[:, i])) + for i in range(X_train_cat.shape[1]) + ] + ) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/synthcity.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/synthcity.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e55c2ee98bfa6beaecf6ad75c7a3b5b2ff05c21e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/synthcity.yaml @@ -0,0 +1,11 @@ +name: synthcity +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pip + - pip: + - synthcity + - category_encoders diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff.yaml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b199bee554e650726bdf7c6b7ac5e52c423cabcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff.yaml @@ -0,0 +1,35 @@ +name: tabdiff +channels: + - pytorch + - nvidia + - defaults +dependencies: + - python=3.10 + - pytorch=2.0.1 + - torchvision==0.15.2 + - torchaudio==2.0.2 + - pytorch-cuda=11.7 + - numpy<2 + - pip + - pip: + - pandas + - scikit-learn + - scipy + - icecream + - xlrd + - tomli-w + - tomli==2.0.1 + - category_encoders + - imbalanced-learn + - kaggle + - transformers + - datasets + - peft==0.9.0 + - ml_collections + - sdmetrics + - prdc + - rdt + - openpyxl + - xgboost + - wandb==0.17.3 + - kaleido diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml new file mode 100644 index 0000000000000000000000000000000000000000..f68214f8b8b4b9b85626bf0ffc743a05d56da67e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/configs/tabdiff_configs.toml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38eff74ecb572c9baafc5e912f8b5421cae686ddb4df6e3bc2a9b69c6ec69cd6 +size 1234 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/main.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/main.py new file mode 100644 index 0000000000000000000000000000000000000000..66e105457a17a8340fde116d0971a1fe05d3479d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/main.py @@ -0,0 +1,344 @@ +import glob +import json +import os +import pickle +import random + +import numpy as np +from tabdiff.metrics import TabMetrics +from tabdiff.modules.main_modules import UniModMLP +from tabdiff.modules.main_modules import Model +from tabdiff.models.unified_ctime_diffusion import UnifiedCtimeDiffusion +from tabdiff.trainer import Trainer +import src +import torch + +from torch.utils.data import DataLoader +import argparse +import warnings + +import wandb + +from copy import deepcopy + +from utils_train import TabDiffDataset + +warnings.filterwarnings('ignore') + + +def main(args): + device = args.device + + ## Disable scientific numerical format + np.set_printoptions(suppress=True) + torch.set_printoptions(sci_mode=False) + + ## Get data info + dataname = args.dataname + data_dir = f'data/{dataname}' + info_path = f'data/{dataname}/info.json' + with open(info_path, 'r') as f: + info = json.load(f) + + ## Set up flags + is_dcr = 'dcr' in dataname + + ## Set experiment name + exp_name = args.exp_name + if args.exp_name is None: + exp_name = 'non_learnable_schedule' if args.non_learnable_schedule else 'learnable_schedule' + exp_name += '_y_only' if args.y_only else '' + + ## Load configs + curr_dir = os.path.dirname(os.path.abspath(__file__)) + config_path = f'{curr_dir}/configs/tabdiff_configs.toml' + raw_config = src.load_config(config_path) + + print(f"{args.mode.capitalize()} Mode is Enabled") + num_samples_to_generate = None + ckpt_path = None + if args.mode == 'train': + print("NEW training is started") + elif args.mode == 'test': + num_samples_to_generate = args.num_samples_to_generate + ckpt_path = args.ckpt_path + if ckpt_path is None: + ckpt_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}" + ckpt_path_arr = glob.glob(f"{ckpt_parent_path}/best_ema_model*") + assert ckpt_path_arr, f"Cannot not infer ckpt_path from {ckpt_parent_path}, please make sure that you first train a model before testing!" + ckpt_path = ckpt_path_arr[0] + config_path = os.path.join(os.path.dirname(ckpt_path), 'config.pkl') + if os.path.exists(config_path): + with open(config_path, 'rb') as f: + cached_raw_config = pickle.load(f) + print(f"Found cached config at {config_path}") + raw_config = cached_raw_config + + + ## Creat model_save and result paths + model_save_path, result_save_path = None, None + if args.mode == 'train': + model_save_path = 'debug/ckpt' if args.debug else f'{curr_dir}/ckpt/{dataname}/{exp_name}' + result_save_path = model_save_path.replace('ckpt', 'result') #i.e., f'{curr_dir}/results/{dataname}/{exp_name}' + elif args.mode == 'test': + if args.report: + result_save_path = f"eval/report_runs/{exp_name}/{dataname}" + else: + result_save_path = os.path.dirname(ckpt_path).replace('ckpt', 'result') # infer the exp_name from the ckpt_name + raw_config['model_save_path'] = model_save_path + raw_config['result_save_path'] = result_save_path + if model_save_path is not None: + if not os.path.exists(model_save_path): + os.makedirs(model_save_path) + if result_save_path is not None: + if not os.path.exists(result_save_path): + os.makedirs(result_save_path) + + ## Make everything determinstic if needed + raw_config['deterministic'] = args.deterministic + if args.deterministic: + print("DETERMINISTIC MODE is enabled!!!") + ## Set global random seeds + torch.manual_seed(0) + random.seed(0) + np.random.seed(0) + + ## Ensure deterministic CUDA operations + os.environ['PYTHONHASHSEED'] = '0' + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' # or ':16:8' + torch.use_deterministic_algorithms(True) + if torch.cuda.is_available(): + torch.cuda.manual_seed(0) + torch.cuda.manual_seed_all(0) + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + ## Set debug mode parameters + if args.debug: # fast eval for DEBUG mode + raw_config['train']['main']['check_val_every'] = 2 + raw_config['diffusion_params']['num_timesteps'] = 4 + raw_config['train']['main']['batch_size'] = 4096 + raw_config['sample']['batch_size'] = 10000 + + # CI /镜像冒烟：覆盖训练步数（默认不设置） + _smoke_steps = os.environ.get("TABDIFF_SMOKE_STEPS", "").strip() + if _smoke_steps and args.mode == "train": + n = max(1, int(_smoke_steps)) + raw_config["train"]["main"]["steps"] = n + raw_config["train"]["main"]["check_val_every"] = max(1, min(n, raw_config["train"]["main"]["check_val_every"])) + # Pipeline 适配器：避免小步数训练时在中途频繁做生成评测；仅在最后一轮 checkpoint + if os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip() and args.mode == "train": + raw_config["train"]["main"]["check_val_every"] = int(raw_config["train"]["main"]["steps"]) + + _train_batch = os.environ.get("TABDIFF_BATCH_SIZE", "").strip() or os.environ.get("TABDIFF_TRAIN_BATCH_SIZE", "").strip() + if _train_batch: + raw_config["train"]["main"]["batch_size"] = max(1, int(_train_batch)) + _sample_batch = os.environ.get("TABDIFF_SAMPLE_BATCH_SIZE", "").strip() + if _sample_batch: + raw_config["sample"]["batch_size"] = max(1, int(_sample_batch)) + _train_lr = os.environ.get("TABDIFF_LR", "").strip() or os.environ.get("TABDIFF_LEARNING_RATE", "").strip() + if _train_lr: + raw_config["train"]["main"]["lr"] = float(_train_lr) + _num_timesteps = os.environ.get("TABDIFF_NUM_TIMESTEPS", "").strip() or os.environ.get("TABDIFF_TIMESTEPS", "").strip() + if _num_timesteps: + raw_config["diffusion_params"]["num_timesteps"] = max(1, int(_num_timesteps)) + + ## Load training data + batch_size = raw_config['train']['main']['batch_size'] + + train_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=True, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + train_loader = DataLoader( + train_data, + batch_size = batch_size, + shuffle = True, + num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')), + ) + d_numerical, categories = train_data.d_numerical, train_data.categories + + val_data = TabDiffDataset(dataname, data_dir, info, y_only=args.y_only, isTrain=False, dequant_dist=raw_config['data']['dequant_dist'], int_dequant_factor=raw_config['data']['int_dequant_factor']) + + ## Load Metrics + real_data_path = f'synthetic/{dataname}/real.csv' + test_data_path = f'synthetic/{dataname}/test.csv' + val_data_path = f'synthetic/{dataname}/val.csv' + if not os.path.exists(val_data_path): + print(f"{args.dataname} does not have its validation set. During MLE evaluation, a validation set will be splitted from the training set!") + val_data_path = None + if args.mode == 'train': + metric_list = ["density"] + else: + if is_dcr: + metric_list = ["dcr"] + else: + metric_list = [ + "density", + "mle", + "c2st", + ] + metrics = TabMetrics(real_data_path, test_data_path, val_data_path, info, device, metric_list=metric_list) + + ## Load the module and models + raw_config['unimodmlp_params']['d_numerical'] = d_numerical + raw_config['unimodmlp_params']['categories'] = (categories+1).tolist() # add one for the mask category + if args.y_only: + raw_config['unimodmlp_params']['use_mlp'] = False # drop the mlp when training the unconditional model + raw_config['unimodmlp_params']['dim_t'] = 128 #reduce the size of the mlp + main_model_path = args.ckpt_path + if main_model_path is None: + main_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name.replace('_y_only', '')}" + main_model_path_arr = glob.glob(f"{main_model_parent_path}/best_ema_model*") + assert main_model_path_arr, f"Cannot not infer the main model's ckpt_path from {main_model_parent_path}, please make sure that you first train a main model before training the y_only model!" + main_model_path = main_model_path_arr[0] + main_model_configs = pickle.load(open(os.path.join(os.path.dirname(main_model_path), 'config.pkl'), 'rb')) + if main_model_configs['diffusion_params']['scheduler'] == "power_mean_per_column": # if learnable schedule is enabled in the main model, we need to infer noise params of the target column from the main model ckpt and train the y_only model with those params + from tabdiff.models.noise_schedule import PowerMeanNoise_PerColumn, LogLinearNoise_PerColumn + if info['task_type'] == 'regression': + noise_schedule = PowerMeanNoise_PerColumn( + num_numerical=main_model_configs['unimodmlp_params']['d_numerical'], + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['rho'] = noise_schedule.rho()[0].item() # the target col is placed at the first position + else: + noise_schedule = LogLinearNoise_PerColumn( + num_categories=len(main_model_configs['unimodmlp_params']['categories']), + **main_model_configs['diffusion_params']['noise_schedule_params'] + ) + raw_config['diffusion_params']['noise_schedule_params']['k'] = noise_schedule.k()[0].item() # the target col is placed at the first position + + backbone = UniModMLP( + **raw_config['unimodmlp_params'] + ) + model = Model(backbone, **raw_config['diffusion_params']['edm_params']) + model.to(device) + + ## Create and load y_only_model for imputation + y_only_model = None + if args.impute: + y_only_model_path = args.y_only_model_path + if y_only_model_path is None: + y_only_model_parent_path = f"{curr_dir}/ckpt/{dataname}/{exp_name}_y_only" + y_only_model_path_arr = glob.glob(f"{y_only_model_parent_path}/best_ema_model*") + assert y_only_model_path_arr, f"Cannot not infer y_only model's ckpt_path from {y_only_model_parent_path}, please make sure that you first train a y_only model before testing imputation!" + y_only_model_path = y_only_model_path_arr[0] + y_only_model_config_path = os.path.join(os.path.dirname(y_only_model_path), 'config.pkl') + with open(y_only_model_config_path, 'rb') as f: + y_only_model_config = pickle.load(f) + y_only_model = UniModMLP( + **y_only_model_config['unimodmlp_params'] + ) + y_only_model = Model(y_only_model, **y_only_model_config['diffusion_params']['edm_params']) + y_only_model.to(device) + # load weights + state_dicts = torch.load(y_only_model_path, map_location=device) + y_only_model.load_state_dict(state_dicts['denoise_fn']) + + if not args.y_only and not args.non_learnable_schedule: + raw_config['diffusion_params']['scheduler'] = 'power_mean_per_column' + raw_config['diffusion_params']['cat_scheduler'] = 'log_linear_per_column' + diffusion = UnifiedCtimeDiffusion( + num_classes=categories, + num_numerical_features=d_numerical, + denoise_fn=model, + y_only_model=y_only_model, + **raw_config['diffusion_params'], + device=device, + ) + num_params = sum(p.numel() for p in diffusion.parameters()) + print("The number of parameters = ", num_params) + diffusion.to(device) + diffusion.train() + + ## Print the configs + printed_configs = json.dumps(raw_config, default=lambda x: int(x) if isinstance(x, np.int64) else x, indent=4) + print(f"The config of the current run is : \n {printed_configs}") + + ## Enable Wandb + project_name = f"tabdiff_{dataname}" + raw_config['project_name'] = project_name + logger = wandb.init( + project=raw_config['project_name'], + name=exp_name, + config=raw_config, + mode='disabled' if args.debug or args.no_wandb else 'online', + ) + + ## Load Trainer + sample_batch_size = raw_config['sample']['batch_size'] + trainer = Trainer( + diffusion, + train_loader, + train_data, + val_data, + metrics, + logger, + **raw_config['train']['main'], + sample_batch_size=sample_batch_size, + num_samples_to_generate=num_samples_to_generate, + model_save_path=raw_config['model_save_path'], + result_save_path=raw_config['result_save_path'], + device=device, + ckpt_path=ckpt_path, + y_only=args.y_only + ) + if args.mode == 'test': + if args.report: + if is_dcr: + trainer.report_test_dcr(args.num_runs) + else: + trainer.report_test(args.num_runs) + elif args.impute: + imputed_sample_save_dir = f"impute/{dataname}/{exp_name}" + trainer.test_impute( + args.trial_start, args.trial_size, + args.resample_rounds, + args.impute_condition, + imputed_sample_save_dir, + args.w_num, + args.w_cat, + ) + else: + trainer.test() + else: + ## Save config + config_save_path = raw_config['model_save_path'] + with open (os.path.join(config_save_path, 'config.pkl'), 'wb') as f: + pickle.dump(raw_config, f) + trainer.run_loop() + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='Training of TabDiff') + + parser.add_argument('--dataname', type=str, default='adult', help='Name of dataset.') + parser.add_argument('--gpu', type=int, default=0, help='GPU index.') + parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) + parser.add_argument('--debug', action='store_true') + parser.add_argument('--no_wandb', action='store_true') + parser.add_argument('--deterministic', action='store_true') + parser.add_argument('--exp_name', type=str, default=None) + parser.add_argument('--non_learnable_schedule', action='store_true') + parser.add_argument('--y_only', action='store_true') + parser.add_argument('--ckpt_path', type=str, default=None) + parser.add_argument('--num_samples_to_generate', type=int, default=None) + parser.add_argument('--report', action='store_true') + parser.add_argument('--num_runs', type=int, default=20) + parser.add_argument('--impute', action='store_true') + parser.add_argument('--trial_start', type=int, default=0) + parser.add_argument('--trial_size', type=int, default=100) + parser.add_argument('--resample_rounds', type=int, default=1) + parser.add_argument('--impute_condition', type=str, default='') + parser.add_argument('--w_num', type=float, default=1.0) + parser.add_argument('--w_cat', type=float, default=1.0) + parser.add_argument('--y_only_model_path', type=str, default=None) + + args = parser.parse_args() + + # check cuda + if args.gpu != -1 and torch.cuda.is_available(): + args.device = f'cuda:{args.gpu}' + else: + args.device = 'cpu' + + main(args) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/metrics.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..4f104427533ac37610c92216b135fdeba3d181be --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/metrics.py @@ -0,0 +1,306 @@ +from copy import deepcopy +import numpy as np +import torch +import pandas as pd +# Metrics +from eval.mle.mle import get_evaluator +from eval.visualize_density import plot_density +from sdmetrics.reports.single_table import QualityReport, DiagnosticReport +from sdmetrics.single_table import LogisticDetection +from sklearn.preprocessing import OneHotEncoder + +from tqdm import tqdm + + +class TabMetrics(object): + def __init__(self, real_data_path, test_data_path, val_data_path, info, device, metric_list) -> None: + self.real_data_path = real_data_path + self.test_data_path = test_data_path + self.val_data_path = val_data_path + self.info = info + self.device = device + self.real_data_size = len(pd.read_csv(real_data_path)) + self.metric_list = metric_list + + def evaluate(self, syn_data): + metrics, extras = {}, {} + syn_data_cp = deepcopy(syn_data) + for metric in self.metric_list: + func = eval(f"self.evaluate_{metric}") + print(f"Evaluating {metric}") + out_metrics, out_extras = func(syn_data_cp) + metrics.update(out_metrics) + extras.update(out_extras) + return metrics, extras + + def evaluate_density(self, syn_data): + real_data = pd.read_csv(self.real_data_path) + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + + info = deepcopy(self.info) + + y_only = len(syn_data.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + syn_data = self.complete_y_only_data(syn_data, real_data, target_col_idx) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} # ensure that keys are all integers? + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + qual_report = QualityReport() + qual_report.generate(new_real_data, new_syn_data, metadata) + + diag_report = DiagnosticReport() + diag_report.generate(new_real_data, new_syn_data, metadata) + + quality = qual_report.get_properties() + diag = diag_report.get_properties() + + Shape = quality['Score'][0] + Trend = quality['Score'][1] + + Overall = (Shape + Trend) / 2 + + shape_details = qual_report.get_details(property_name='Column Shapes') + trend_details = qual_report.get_details(property_name='Column Pair Trends') + + if y_only: + Shape = shape_details['Score'].min() + out_metrics = { + "density/Shape": Shape, + "density/Trend": Trend, + "density/Overall": Overall, + } + out_extras = { + "shapes": shape_details, + "trends": trend_details + } + return out_metrics, out_extras + + def evaluate_mle(self, syn_data): + train = syn_data.to_numpy() + test = pd.read_csv(self.test_data_path).to_numpy() + val = pd.read_csv(self.val_data_path).to_numpy() if self.val_data_path else None + + info = deepcopy(self.info) + + task_type = info['task_type'] + + evaluator = get_evaluator(task_type) + + if task_type == 'regression': + best_r2_scores, best_rmse_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_r2_scores', 'best_rmse_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + else: + best_f1_scores, best_weighted_scores, best_auroc_scores, best_acc_scores, best_avg_scores = evaluator(train, test, info, val=val) + + overall_scores = {} + for score_name in ['best_f1_scores', 'best_weighted_scores', 'best_auroc_scores', 'best_acc_scores', 'best_avg_scores']: + overall_scores[score_name] = {} + + scores = eval(score_name) + for method in scores: + name = method['name'] + method.pop('name') + overall_scores[score_name][name] = method + + mle_score = overall_scores['best_rmse_scores']['XGBRegressor']['RMSE'] if task_type == 'regression' else overall_scores['best_auroc_scores']['XGBClassifier']['roc_auc'] + out_metrics = { + "mle": mle_score, + } + out_extras = { + "mle": overall_scores, + } + return out_metrics, out_extras + + def evaluate_c2st(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + + real_data.columns = range(len(real_data.columns)) + syn_data.columns = range(len(syn_data.columns)) + + metadata = info['metadata'] + metadata['columns'] = {int(key): value for key, value in metadata['columns'].items()} + + new_real_data, new_syn_data, metadata = reorder(real_data, syn_data, info) + + score = LogisticDetection.compute( + real_data=new_real_data, + synthetic_data=new_syn_data, + metadata=metadata + ) + + out_metrics = { + "c2st": score, + } + out_extras = {} + return out_metrics, out_extras + + def evaluate_dcr(self, syn_data): + info = deepcopy(self.info) + real_data = pd.read_csv(self.real_data_path) + test_data = pd.read_csv(self.test_data_path) + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + num_ranges = [] + + real_data.columns = list(np.arange(len(real_data.columns))) + syn_data.columns = list(np.arange(len(real_data.columns))) + test_data.columns = list(np.arange(len(real_data.columns))) + for i in num_col_idx: + num_ranges.append(real_data[i].max() - real_data[i].min()) + + num_ranges = np.array(num_ranges) + + + num_real_data = real_data[num_col_idx] + cat_real_data = real_data[cat_col_idx] + num_syn_data = syn_data[num_col_idx] + cat_syn_data = syn_data[cat_col_idx] + num_test_data = test_data[num_col_idx] + cat_test_data = test_data[cat_col_idx] + + num_real_data_np = num_real_data.to_numpy() + cat_real_data_np = cat_real_data.to_numpy().astype('str') + num_syn_data_np = num_syn_data.to_numpy() + cat_syn_data_np = cat_syn_data.to_numpy().astype('str') + num_test_data_np = num_test_data.to_numpy() + cat_test_data_np = cat_test_data.to_numpy().astype('str') + + encoder = OneHotEncoder() + cat_complete_data_np = np.concatenate([cat_real_data_np, cat_test_data_np], axis=0) + encoder.fit(cat_complete_data_np) + # encoder.fit(cat_real_data_np) + + + cat_real_data_oh = encoder.transform(cat_real_data_np).toarray() + cat_syn_data_oh = encoder.transform(cat_syn_data_np).toarray() + cat_test_data_oh = encoder.transform(cat_test_data_np).toarray() + + num_real_data_np = num_real_data_np / num_ranges + num_syn_data_np = num_syn_data_np / num_ranges + num_test_data_np = num_test_data_np / num_ranges + + real_data_np = np.concatenate([num_real_data_np, cat_real_data_oh], axis=1) + syn_data_np = np.concatenate([num_syn_data_np, cat_syn_data_oh], axis=1) + test_data_np = np.concatenate([num_test_data_np, cat_test_data_oh], axis=1) + + device = self.device + + real_data_th = torch.tensor(real_data_np).to(device) + syn_data_th = torch.tensor(syn_data_np).to(device) + test_data_th = torch.tensor(test_data_np).to(device) + + dcrs_real = [] + dcrs_test = [] + batch_size = 10000 // cat_real_data_oh.shape[1] # This esitmation should make sure that dcr_real and dcr_test can be fit into 10GB GPU memory + + for i in tqdm(range((syn_data_th.shape[0] // batch_size) + 1)): + if i != (syn_data_th.shape[0] // batch_size): + batch_syn_data_th = syn_data_th[i*batch_size: (i+1) * batch_size] + else: + batch_syn_data_th = syn_data_th[i*batch_size:] + + dcr_real = (batch_syn_data_th[:, None] - real_data_th).abs().sum(dim = 2).min(dim = 1).values + dcr_test = (batch_syn_data_th[:, None] - test_data_th).abs().sum(dim = 2).min(dim = 1).values + dcrs_real.append(dcr_real) + dcrs_test.append(dcr_test) + + dcrs_real = torch.cat(dcrs_real) + dcrs_test = torch.cat(dcrs_test) + + score = (dcrs_real < dcrs_test).nonzero().shape[0] / dcrs_real.shape[0] + + out_metrics = { + "dcr": score, + } + out_extras = { + "dcr_real": dcrs_real.cpu().numpy(), + "dcr_test": dcrs_test.cpu().numpy(), + } + return out_metrics, out_extras + + + def plot_density(self, syn_data): + syn_data_cp = deepcopy(syn_data) + real_data = pd.read_csv(self.real_data_path) + info = deepcopy(self.info) + y_only = len(syn_data_cp.columns)==1 + if y_only: + target_col_idx = info['target_col_idx'][0] + target_col_name = info['column_names'][target_col_idx] + syn_data_cp = self.complete_y_only_data(syn_data_cp, real_data, target_col_name) + img = plot_density(syn_data_cp, real_data, info) + return img + + def complete_y_only_data(self, syn_data, real_data, target_col_idx): + syn_target_col = deepcopy(syn_data.iloc[:, 0]) + syn_data = deepcopy(real_data) + syn_data[target_col_idx] = syn_target_col + return syn_data + + +def reorder(real_data, syn_data, info): + num_col_idx = deepcopy(info['num_col_idx']) # BUG: info will be modified by += in the next few lines + cat_col_idx = deepcopy(info['cat_col_idx']) + target_col_idx = deepcopy(info['target_col_idx']) + + task_type = info['task_type'] + if task_type == 'regression': + num_col_idx += target_col_idx + else: + cat_col_idx += target_col_idx + + real_num_data = real_data[num_col_idx] + real_cat_data = real_data[cat_col_idx] + + new_real_data = pd.concat([real_num_data, real_cat_data], axis=1) + new_real_data.columns = range(len(new_real_data.columns)) + + syn_num_data = syn_data[num_col_idx] + syn_cat_data = syn_data[cat_col_idx] + + new_syn_data = pd.concat([syn_num_data, syn_cat_data], axis=1) + new_syn_data.columns = range(len(new_syn_data.columns)) + + + metadata = info['metadata'] + + columns = metadata['columns'] + metadata['columns'] = {} + + inverse_idx_mapping = info['inverse_idx_mapping'] + + + for i in range(len(new_real_data.columns)): + if i < len(num_col_idx): + metadata['columns'][i] = columns[num_col_idx[i]] + else: + metadata['columns'][i] = columns[cat_col_idx[i-len(num_col_idx)]] + + + return new_real_data, new_syn_data, metadata \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/models/noise_schedule.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/models/noise_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0fc02778162edbf35ef9574aad67483c9e6a7cf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/models/noise_schedule.py @@ -0,0 +1,157 @@ +import abc + +import torch +import torch.nn as nn + + +class Noise(abc.ABC, nn.Module): + """ + Baseline forward method to get the total + rate of noise at a timestep + """ + def forward(self, t): + # Assume time goes from 0 to 1 + return self.total_noise(t), self.rate_noise(t) + + @abc.abstractmethod + def total_noise(self, t): + """ + Total noise ie \int_0^t g(t) dt + g(0) + """ + pass + + +class LogLinearNoise(Noise): + """Log Linear noise schedule. + + """ + def __init__(self, eps_max=1e-3, eps_min=1e-5, **kwargs): + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + self.sigma_max = self.total_noise(torch.tensor(1.0)) + self.sigma_min = self.total_noise(torch.tensor(0.0)) + + def k(self): + return torch.tensor(1) + + def rate_noise(self, t): + return (1 - self.eps_max - self.eps_min) / (1 - ((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + + def total_noise(self, t): + """ + sigma_min=-log(1-eps_min), when t=0 + sigma_max=-log(eps_max), when t=1 + """ + return -torch.log1p(-((1 - self.eps_max - self.eps_min) * t + self.eps_min)) + +class PowerMeanNoise(Noise): + """The noise schedule using the power mean interpolation function. + + This is the schedule used in EDM + """ + def __init__(self, sigma_min=0.002, sigma_max=80, rho=7, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.raw_rho = rho + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return torch.tensor(self.raw_rho) + + def total_noise(self, t): + sigma = (self.sigma_min ** (1/self.rho()) + t * ( + self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho()))).pow(self.rho()) + return sigma + + def inverse_to_t(self, sigma): + t = (sigma.pow(1/self.rho()) - self.sigma_min ** (1/self.rho())) / (self.sigma_max ** (1/self.rho()) - self.sigma_min ** (1/self.rho())) + return t + + +class PowerMeanNoise_PerColumn(nn.Module): + + def __init__(self, num_numerical, sigma_min=0.002, sigma_max=80, rho_init=1, rho_offset=2, **kwargs): + super().__init__() + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.num_numerical = num_numerical + self.rho_offset = rho_offset + self.rho_raw = nn.Parameter(torch.tensor([rho_init] * self.num_numerical, dtype=torch.float32)) + + def rho(self): + # Return the softplus-transformed rho for all num_numerical values + return nn.functional.softplus(self.rho_raw) + self.rho_offset + + def total_noise(self, t): + """ + Compute total noise for each t in the batch for all num_numerical rhos. + t: [batch_size] + Returns: [batch_size, num_numerical] + """ + batch_size = t.size(0) + + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + sigma = (sigma_min_pow + t * (sigma_max_pow - sigma_min_pow)).pow(rho) # Shape: [batch_size, num_numerical] + + return sigma + + def rate_noise(self, t): + return None + + def inverse_to_t(self, sigma): + """ + Inverse function to map sigma back to t, with proper broadcasting support. + sigma: [batch_size, num_numerical] or [batch_size, 1] + Returns: t: [batch_size, num_numerical] + """ + rho = self.rho() + + sigma_min_pow = self.sigma_min ** (1 / rho) # Shape: [num_numerical] + sigma_max_pow = self.sigma_max ** (1 / rho) # Shape: [num_numerical] + + # To enable broadcasting between sigma and the per-column rho values, expand rho where needed. + t = (sigma.pow(1 / rho) - sigma_min_pow) / (sigma_max_pow - sigma_min_pow) + + return t + + +class LogLinearNoise_PerColumn(nn.Module): + + def __init__(self, num_categories, eps_max=1e-3, eps_min=1e-5, k_init=-6, k_offset=1, **kwargs): + + super().__init__() + self.eps_max = eps_max + self.eps_min = eps_min + # Use softplus to ensure k is positive + self.num_categories = num_categories + self.k_offset = k_offset + self.k_raw = nn.Parameter(torch.tensor([k_init] * self.num_categories, dtype=torch.float32)) + + def k(self): + return torch.nn.functional.softplus(self.k_raw) + self.k_offset + + def rate_noise(self, t, noise_fn=None): + """ + Compute rate noise for all categories with broadcasting. + t: [batch_size] + Returns: [batch_size, num_categories] + """ + k = self.k() # Shape: [num_categories] + + numerator = (1 - self.eps_max - self.eps_min) * k * t.pow(k - 1) + denominator = 1 - ((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min) + rate = numerator / denominator # Shape: [batch_size, num_categories] + + return rate + + def total_noise(self, t, noise_fn=None): + k = self.k() # Shape: [num_categories] + + total_noise = -torch.log1p(-((1 - self.eps_max - self.eps_min) * t.pow(k) + self.eps_min)) + + return total_noise diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..599463b6cdf2528c5f35ebd5ddc5c511e5f34d8f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/models/unified_ctime_diffusion.py @@ -0,0 +1,597 @@ +import torch.nn.functional as F +import torch +import math +import numpy as np +from tabdiff.models.noise_schedule import * +from tqdm import tqdm +from itertools import chain + +""" +“Our implementation of the continuous-time masked diffusion is inspired by https://arxiv.org/abs/2406.07524's implementation at [https://github.com/kuleshov-group/mdlm], with modifications to support data distributions that include categorical dimensions of different sizes.” +""" + +S_churn= 1 +S_min=0 +S_max=float('inf') +S_noise=1 + +class UnifiedCtimeDiffusion(torch.nn.Module): + def __init__( + self, + num_classes: np.array, + num_numerical_features: int, + denoise_fn, + y_only_model, + num_timesteps=1000, + scheduler='power_mean', + cat_scheduler='log_linear', + noise_dist='uniform', + edm_params={}, + noise_dist_params={}, + noise_schedule_params={}, + sampler_params={}, + device=torch.device('cpu'), + **kwargs + ): + + super(UnifiedCtimeDiffusion, self).__init__() + + self.num_numerical_features = num_numerical_features + self.num_classes = num_classes # it as a vector [K1, K2, ..., Km] + self.num_classes_expanded = torch.from_numpy( + np.concatenate([num_classes[i].repeat(num_classes[i]) for i in range(len(num_classes))]) + ).to(device) if len(num_classes)>0 else torch.tensor([]).to(device).int() + self.mask_index = torch.tensor(self.num_classes).long().to(device) + self.neg_infinity = -1000000.0 + self.num_classes_w_mask = tuple(self.num_classes + 1) + + offsets = np.cumsum(self.num_classes) + offsets = np.append([0], offsets) + self.slices_for_classes = [] + for i in range(1, len(offsets)): + self.slices_for_classes.append(np.arange(offsets[i - 1], offsets[i])) + self.offsets = torch.from_numpy(offsets).to(device) + + offsets = np.cumsum(self.num_classes) + np.arange(1, len(self.num_classes)+1) + offsets = np.append([0], offsets) + self.slices_for_classes_with_mask = [] + for i in range(1, len(offsets)): + self.slices_for_classes_with_mask.append(np.arange(offsets[i - 1], offsets[i])) + + self._denoise_fn = denoise_fn + self.y_only_model = y_only_model + self.num_timesteps = num_timesteps + self.scheduler = scheduler + self.cat_scheduler = cat_scheduler + self.noise_dist = noise_dist + self.edm_params = edm_params + self.noise_dist_params = noise_dist_params + self.sampler_params = sampler_params + if self.num_numerical_features == 0: + self.sampler_params['stochastic_sampler'] = False + self.sampler_params['second_order_correction'] = False + + self.w_num = 0.0 + self.w_cat = 0.0 + self.num_mask_idx = [] + self.cat_mask_idx = [] + + self.device = device + + if self.scheduler == 'power_mean': + self.num_schedule = PowerMeanNoise(**noise_schedule_params) + elif self.scheduler == 'power_mean_per_column': + self.num_schedule = PowerMeanNoise_PerColumn(num_numerical = num_numerical_features, **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.scheduler}-- is not implemented for contiuous data at CTIME ") + + if self.cat_scheduler == 'log_linear': + self.cat_schedule = LogLinearNoise(**noise_schedule_params) + elif self.cat_scheduler == 'log_linear_per_column': + self.cat_schedule = LogLinearNoise_PerColumn(num_categories = len(num_classes), **noise_schedule_params) + else: + raise NotImplementedError(f"The noise schedule--{self.cat_scheduler}-- is not implemented for discrete data at CTIME ") + + def mixed_loss(self, x): + b = x.shape[0] + device = x.device + + x_num = x[:, :self.num_numerical_features] + x_cat = x[:, self.num_numerical_features:].long() + # Sample noise level + if self.noise_dist == "uniform_t": + t = torch.rand(b, device=device, dtype=x_num.dtype) + t = t[:, None] + sigma_num = self.num_schedule.total_noise(t) + sigma_cat = self.cat_schedule.total_noise(t) + dsigma_cat = self.cat_schedule.rate_noise(t) + else: + sigma_num = self.sample_ctime_noise(x) + t = self.num_schedule.inverse_to_t(sigma_num) + while torch.any((t < 0) + (t > 1)): + # restrict t to [0,1] + # this iterative approach is equivalent to sampling from a truncated version of the orignal noise distribution + invalid_idx = ((t < 0) + (t > 1)).nonzero().squeeze(-1) + sigma_num[invalid_idx] = self.sample_ctime_noise(x[:len(invalid_idx)]) + t = self.num_schedule.inverse_to_t(sigma_num) + assert not torch.any((t < 0) + (t > 1)) + sigma_cat = self.cat_schedule.total_noise(t) + # Convert sigma_cat to the corresponding alpha and move_chance + alpha = torch.exp(-sigma_cat) + move_chance = -torch.expm1(-sigma_cat) # torch.expm1 gives better numertical stability + + # Continuous forward diff + x_num_t = x_num + if x_num.shape[1] > 0: + noise = torch.randn_like(x_num) + x_num_t = x_num + noise * sigma_num + + # Discrete forward diff + x_cat_t = x_cat + x_cat_t_soft = x_cat # in the case where x_cat is empty, x_cat_t_soft will have the same shape as x_cat + if x_cat.shape[1] > 0: + is_learnable = self.cat_scheduler == 'log_linear_per_column' + strategy = 'soft'if is_learnable else 'hard' + x_cat_t, x_cat_t_soft = self.q_xt(x_cat, move_chance, strategy=strategy) + + # Predict orignal data (distribution) + model_out_num, model_out_cat = self._denoise_fn( + x_num_t, x_cat_t_soft, + t.squeeze(), sigma=sigma_num + ) + + d_loss = torch.zeros((1,)).float() + c_loss = torch.zeros((1,)).float() + + if x_num.shape[1] > 0: + c_loss = self._edm_loss(model_out_num, x_num, sigma_num) + if x_cat.shape[1] > 0: + logits = self._subs_parameterization(model_out_cat, x_cat_t) # log normalized probabilities, with the entry mask category being set to -inf + d_loss = self._absorbed_closs(logits, x_cat, sigma_cat, dsigma_cat) + + return d_loss.mean(), c_loss.mean() + + @torch.no_grad() + def sample(self, num_samples): + b = num_samples + device = self.device + dtype = torch.float32 + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + z_norm = torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + z_norm, z_cat, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + + assert torch.all(z_cat < self.mask_index) + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + + def sample_all(self, num_samples, batch_size, keep_nan_samples=False): + b = batch_size + + all_samples = [] + num_generated = 0 + while num_generated < num_samples: + print(f"Samples left to generate: {num_samples-num_generated}") + sample = self.sample(b) + mask_nan = torch.any(sample.isnan(), dim=1) + if keep_nan_samples: + # If the sample instances that contains Nan are decided to be kept, the row with Nan will be foreced to all zeros + sample = sample * (~mask_nan)[:, None] + else: + # Otherwise the instances with Nan will be eliminated + sample = sample[~mask_nan] + + all_samples.append(sample) + num_generated += sample.shape[0] + + x_gen = torch.cat(all_samples, dim=0)[:num_samples] + + return x_gen + + def q_xt(self, x, move_chance, strategy='hard'): + """Computes the noisy sample xt. + + Args: + x: int torch.Tensor with shape (batch_size, + diffusion_model_input_length), input. + move_chance: float torch.Tensor with shape (batch_size, 1). + """ + if strategy == 'hard': + move_indices = torch.rand( + * x.shape, device=x.device) < move_chance + xt = torch.where(move_indices, self.mask_index, x) + xt_soft = self.to_one_hot(xt).to(move_chance.dtype) + return xt, xt_soft + elif strategy == 'soft': + bs = x.shape[0] + xt_soft = torch.zeros(bs, torch.sum(self.mask_index+1), device=x.device) + xt = torch.zeros_like(x) + for i in range(len(self.num_classes)): + slice_i = self.slices_for_classes_with_mask[i] + # set the bernoulli probabilities, which determines the "coin flip" transition to the mask class + prob_i = torch.zeros(bs, 2, device=x.device) + prob_i[:,0] = 1-move_chance[:,i] + prob_i[:,-1] = move_chance[:,i] + log_prob_i = torch.log(prob_i) + # draw soft samples and place them back to the corresponding columns + soft_sample_i = F.gumbel_softmax(log_prob_i, tau=0.01, hard=True) + idx = torch.stack((x[:,i]+slice_i[0], torch.ones_like(x[:,i])*slice_i[-1]), dim=-1) + xt_soft[torch.arange(len(idx)).unsqueeze(1), idx] = soft_sample_i + # retrieve the hard samples + xt[:, i] = torch.where(soft_sample_i[:,1] > soft_sample_i[:,0], self.mask_index[i], x[:,i]) + return xt, xt_soft + + + def _subs_parameterization(self, unormalized_prob, xt): + # log prob at the mask index = - infinity + unormalized_prob = self.pad(unormalized_prob, self.neg_infinity) + + unormalized_prob[:, range(unormalized_prob.shape[1]), self.mask_index] += self.neg_infinity + + # Take log softmax on the unnormalized probabilities to the logits + logits = unormalized_prob - torch.logsumexp(unormalized_prob, dim=-1, + keepdim=True) + # Apply updates directly in the logits matrix. + # For the logits of the unmasked tokens, set all values + # to -infinity except for the indices corresponding to + # the unmasked tokens. + unmasked_indices = (xt != self.mask_index) # (bs, K) + logits[unmasked_indices] = self.neg_infinity + logits[unmasked_indices, xt[unmasked_indices]] = 0 + return logits + + def pad(self, x, pad_value): + """ + Converts a concatenated tensor of class probabilities into a padded matrix, + where each sub-tensor is padded along the last dimension to match the largest + category size (max number of classes). + + Args: + x (Tensor): The input tensor containing concatenated probabilities for all the categories in x_cat. + [bs, sum(num_classes_w_mask)] + pad_value (float): The value filled into the dummy entries, which are padded to ensure all sub-tensors have equal size + along the last dimension. + + Returns: + Tensor: A new tensorwith + [bs, len(num_classes_w_mask), max(num_classes_w_mask)), num_categories] + """ + splited = torch.split(x, self.num_classes_w_mask, dim=-1) + max_K = max(self.num_classes_w_mask) + padded_ = [ + torch.cat(( + t, + pad_value*torch.ones(*(t.shape[:-1]), max_K-t.shape[-1], dtype=t.dtype, device=t.device) + ), dim=-1) + for t in splited] + out = torch.stack(padded_, dim=-2) + return out + + def to_one_hot(self, x_cat): + x_cat_oh = torch.cat( + [F.one_hot(x_cat[:, i], num_classes=self.num_classes[i]+1,) for i in range(len(self.num_classes))], + dim=-1 + ) + return x_cat_oh + + def _absorbed_closs(self, model_output, x0, sigma, dsigma): + """ + alpha: (bs,) + """ + log_p_theta = torch.gather( + model_output, -1, x0[:, :, None] + ).squeeze(-1) + alpha = torch.exp(-sigma) + if self.cat_scheduler in ['log_linear_unified', 'log_linear_per_column']: + elbo_weight = - dsigma / torch.expm1(sigma) + else: + elbo_weight = -1/(1-alpha) + + loss = elbo_weight * log_p_theta + return loss + + def _sample_masked_prior(self, *batch_dims): + return self.mask_index[None,:] * torch.ones( + * batch_dims, dtype=torch.int64, device=self.mask_index.device) + + def _mdlm_update(self, log_p_x0, x, alpha_t, alpha_s): + """ + # t: (bs,) + log_p_x0: (bs, K, K_max) + # alpha_t: (bs,) + # alpha_s: (bs,) + alpha_t: (bs, 1/K_cat) + alpha_s: (bs,1/K_cat) + """ + move_chance_t = 1 - alpha_t + move_chance_s = 1 - alpha_s + move_chance_t = move_chance_t.unsqueeze(-1) + move_chance_s = move_chance_s.unsqueeze(-1) + assert move_chance_t.ndim == log_p_x0.ndim + # Technically, this isn't q_xs since there's a division + # term that is missing. This division term doesn't affect + # the samples. + # There is a noremalizing term is (1-\alpha_t) who's responsility is to ensure q_xs is normalized. + # However, omiting it won't make a difference for the Gumbel-max sampling trick in _sample_categorical() + q_xs = log_p_x0.exp() * (move_chance_t + - move_chance_s) + q_xs[:, range(q_xs.shape[1]), self.mask_index] = move_chance_s[:, :, 0] + + # Important: make sure that prob of dummy classes are exactly 0 + dummy_mask = torch.tensor([[(1 if i <= mask_idx else 0) for i in range(max(self.mask_index+1))] for mask_idx in self.mask_index], device=q_xs.device) + dummy_mask = torch.ones_like(q_xs) * dummy_mask + q_xs *= dummy_mask + + _x = self._sample_categorical(q_xs) + + copy_flag = (x != self.mask_index).to(x.dtype) + + z_cat = copy_flag * x + (1 - copy_flag) * _x + return copy_flag * x + (1 - copy_flag) * _x, q_xs + + def _sample_categorical(self, categorical_probs): + gumbel_norm = ( + 1e-10 + - (torch.rand_like(categorical_probs) + 1e-10).log()) + return (categorical_probs / gumbel_norm).argmax(dim=-1) + + def sample_ctime_noise(self, batch): + if self.noise_dist == 'log_norm': + rnd_normal = torch.randn(batch.shape[0], device=batch.device) + sigma = (rnd_normal * self.noise_dist_params['P_std'] + self.noise_dist_params['P_mean']).exp() + else: + raise NotImplementedError(f"The noise distribution--{self.noise_dist}-- is not implemented for CTIME ") + return sigma + + def _edm_loss(self, D_yn, y, sigma): + weight = (sigma ** 2 + self.edm_params['sigma_data'] ** 2) / (sigma * self.edm_params['sigma_data']) ** 2 + + target = y + loss = weight * ((D_yn - target) ** 2) + + return loss + + def edm_update( + self, x_num_cur, x_cat_cur, i, + t_cur, t_next, t_hat, + sigma_num_cur, sigma_num_next, sigma_num_hat, + sigma_cat_cur, sigma_cat_next, sigma_cat_hat, + ): + """ + i = T-1,...,0 + """ + cfg = self.y_only_model is not None + + b = x_num_cur.shape[0] + has_cat = len(self.num_classes) > 0 + + # Get x_num_hat by move towards the noise by a small step + x_num_hat = x_num_cur + (sigma_num_hat ** 2 - sigma_num_cur ** 2).sqrt() * S_noise * torch.randn_like(x_num_cur) + # Get x_cat_hat + move_chance = -torch.expm1(sigma_cat_cur - sigma_cat_hat) # the incremental move change is 1 - alpha_t/alpha_s = 1 - exp(sigma_s - sigma_t) + x_cat_hat, _ = self.q_xt(x_cat_cur, move_chance) if has_cat else (x_cat_cur, x_cat_cur) + + # Get predictions + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_hat.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_hat.float(), x_cat_hat_oh, + t_hat.squeeze().repeat(b), sigma=sigma_num_hat.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + # Apply cfg updates, if is in cfg mode + is_bin_class = len(self.num_mask_idx) == 0 + is_learnable = self.scheduler=="power_mean_per_column" + if cfg: + if not is_learnable: + sigma_cond = sigma_num_hat + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_hat * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_hat[self.num_mask_idx] + y_num_hat = x_num_hat.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:,idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_hat, + y_cat_hat, + t_hat.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + mask_logit_idx = [self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx] + mask_logit_idx = np.concatenate(mask_logit_idx) if len(mask_logit_idx)>0 else np.array([]) + + raw_logits[:, mask_logit_idx] *= 1 + self.w_cat + raw_logits[:, mask_logit_idx] -= self.w_cat*y_only_raw_logits + + # Euler step + d_cur = (x_num_hat - denoised) / sigma_num_hat + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * d_cur + + # Unmasking + x_cat_next = x_cat_cur + q_xs = torch.zeros_like(x_cat_cur).float() + if has_cat: + logits = self._subs_parameterization(raw_logits, x_cat_hat) + alpha_t = torch.exp(-sigma_cat_hat).unsqueeze(0).repeat(b,1) + alpha_s = torch.exp(-sigma_cat_next).unsqueeze(0).repeat(b,1) + x_cat_next, q_xs = self._mdlm_update(logits, x_cat_hat, alpha_t, alpha_s) + + # Apply 2nd order correction. + if self.sampler_params['second_order_correction']: + if i > 0: + x_cat_hat_oh = self.to_one_hot(x_cat_hat).to(x_num_next.dtype) if has_cat else x_cat_hat + denoised, raw_logits = self._denoise_fn( + x_num_next.float(), x_cat_hat_oh, + t_next.squeeze().repeat(b), sigma=sigma_num_next.unsqueeze(0).repeat(b,1) + ) + if cfg: + if not is_learnable: + sigma_cond = sigma_num_next + else: + if is_bin_class: + sigma_cond = (0.002 ** (1/7) + t_next * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma_num_next[self.num_mask_idx] + y_num_next = x_num_next.float()[:, self.num_mask_idx] + idx = list(chain(*[self.slices_for_classes_with_mask[i] for i in self.cat_mask_idx])) + y_cat_hat = x_cat_hat_oh[:, idx] + y_only_denoised, y_only_raw_logits = self.y_only_model( + y_num_next, + y_cat_hat, + t_next.squeeze().repeat(b), sigma=sigma_cond.unsqueeze(0).repeat(b,1) # sigma accepts (bs, K_num) + ) + denoised[:, self.num_mask_idx] *= 1 + self.w_num + denoised[:, self.num_mask_idx] -= self.w_num*y_only_denoised + + d_prime = (x_num_next - denoised) / sigma_num_next + x_num_next = x_num_hat + (sigma_num_next - sigma_num_hat) * (0.5 * d_cur + 0.5 * d_prime) + + return x_num_next, x_cat_next, q_xs + + + def sample_impute(self, x_num, x_cat, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat): + self.w_num = w_num + self.w_cat = w_cat + self.num_mask_idx = num_mask_idx + self.cat_mask_idx = cat_mask_idx + + b = x_num.size(0) + device = self.device + dtype = torch.float32 + + # Create masks, true for the missing columns + num_mask = [i in num_mask_idx for i in range(self.num_numerical_features)] + cat_mask = [i in cat_mask_idx for i in range(len(self.num_classes))] + num_mask = torch.tensor(num_mask).to(x_num.device).to(x_num.dtype) + cat_mask = torch.tensor(cat_mask).to(x_cat.device).to(x_cat.dtype) + + # Create the chain of t + t = torch.linspace(0,1,self.num_timesteps, dtype=dtype, device=device) # times = 0.0,...,1.0 + t = t[:, None] + + # Compute the chains of sigma + sigma_num_cur = self.num_schedule.total_noise(t) + sigma_cat_cur = self.cat_schedule.total_noise(t) + sigma_num_next = torch.zeros_like(sigma_num_cur) + sigma_num_next[1:] = sigma_num_cur[0:-1] + sigma_cat_next = torch.zeros_like(sigma_cat_cur) + sigma_cat_next[1:] = sigma_cat_cur[0:-1] + + # Prepare sigma_hat for stochastic sampling mode + if self.sampler_params['stochastic_sampler']: + gamma = min(S_churn / self.num_timesteps, np.sqrt(2) - 1) * (S_min <= sigma_num_cur) * (sigma_num_cur <= S_max) + sigma_num_hat = sigma_num_cur + gamma * sigma_num_cur + t_hat = self.num_schedule.inverse_to_t(sigma_num_hat) + t_hat = torch.min(t_hat, dim=-1, keepdim=True).values # take the samllest t_hat induced by sigma_num + zero_gamma = (gamma==0).any() + t_hat[zero_gamma] = t[zero_gamma] + out_of_bound = (t_hat > 1).squeeze() + sigma_num_hat[out_of_bound] = sigma_num_cur[out_of_bound] + t_hat[out_of_bound] = t[out_of_bound] + sigma_cat_hat = self.cat_schedule.total_noise(t_hat) + else: + t_hat = t + sigma_num_hat = sigma_num_cur + sigma_cat_hat = sigma_cat_cur + + # Sample priors for the continuous dimensions + if impute_condition == "x_t": + z_norm = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_cur[-1] # z_{t_max} = x_0(masked) + sigma_max*epsilon + elif impute_condition == "x_0": + z_norm = x_num + + # Sample priors for the discrete dimensions + has_cat = len(self.num_classes) > 0 + z_cat = torch.zeros((b, 0), device=device).float() # the default values for categorical sample if the dataset has no categorical entry + if has_cat: + if impute_condition == "x_t": + z_cat = self._sample_masked_prior( + b, + len(self.num_classes), + ) # z_{t_max} is still all pushed to [MASK] + elif impute_condition == "x_0": + z_cat = x_cat + + pbar = tqdm(reversed(range(0, self.num_timesteps)), total=self.num_timesteps) + pbar.set_description(f"Sampling Progress") + for i in pbar: + for u in range (resample_rounds): + # Get known parts by Forward Flow + if impute_condition == "x_t": + z_norm_known = x_num + torch.randn((b, self.num_numerical_features), device=device) * sigma_num_next[i] + move_chance = 1 - torch.exp(-sigma_cat_next[i]) if i < (self.num_timesteps-1) else torch.ones_like(sigma_cat_next[i]) # force move_chance to be 1 for the first iteration + z_cat_known, _ = self.q_xt(x_cat, move_chance) + elif impute_condition == "x_0": + z_norm_known = x_num + z_cat_known = x_cat + + # Get unknown by Reverse Step + z_norm_unknown, z_cat_unknown, q_xs = self.edm_update( + z_norm, z_cat, i, + t[i], t[i-1] if i > 0 else None, t_hat[i], + sigma_num_cur[i], sigma_num_next[i], sigma_num_hat[i], + sigma_cat_cur[i], sigma_cat_next[i], sigma_cat_hat[i], + ) + z_norm = (1 - num_mask) * z_norm_known + num_mask * z_norm_unknown + z_cat = (1 - cat_mask) * z_cat_known + cat_mask * z_cat_unknown + + # Resample x_t from x_{t-1} by Foward Step + if u < resample_rounds-1: + z_norm = z_norm + (sigma_num_cur[i] ** 2 - sigma_num_next[i] ** 2).sqrt() * S_noise * torch.randn_like(z_norm) + move_chance = -torch.expm1(sigma_cat_next[i] - sigma_cat_cur[i]) + z_cat, _ = self.q_xt(z_cat, move_chance) + + sample = torch.cat([z_norm, z_cat], dim=1).cpu() + return sample + \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/modules/main_modules.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/modules/main_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..8e4333d626894e114ab775f07ceba5c8c46b52f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/modules/main_modules.py @@ -0,0 +1,167 @@ +from typing import Callable, Union + +from tabdiff.modules.transformer import Reconstructor, Tokenizer, Transformer +import torch +import torch.nn as nn +import torch.optim + +ModuleType = Union[str, Callable[..., nn.Module]] + +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + +class PositionalEmbedding(torch.nn.Module): + def __init__(self, num_channels, max_positions=10000, endpoint=False): + super().__init__() + self.num_channels = num_channels + self.max_positions = max_positions + self.endpoint = endpoint + + def forward(self, x): + freqs = torch.arange(start=0, end=self.num_channels//2, dtype=torch.float32, device=x.device) + freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0)) + freqs = (1 / self.max_positions) ** freqs + x = x.ger(freqs.to(x.dtype)) + x = torch.cat([x.cos(), x.sin()], dim=1) + return x + + +class MLPDiffusion(nn.Module): + def __init__(self, d_in, dim_t = 512, use_mlp=True): + super().__init__() + self.dim_t = dim_t + + self.proj = nn.Linear(d_in, dim_t) + + self.mlp = nn.Sequential( + nn.Linear(dim_t, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t * 2), + nn.SiLU(), + nn.Linear(dim_t * 2, dim_t), + nn.SiLU(), + nn.Linear(dim_t, d_in), + ) if use_mlp else nn.Linear(dim_t, d_in) + + self.map_noise = PositionalEmbedding(num_channels=dim_t) + self.time_embed = nn.Sequential( + nn.Linear(dim_t, dim_t), + nn.SiLU(), + nn.Linear(dim_t, dim_t) + ) + + self.use_mlp = use_mlp + + def forward(self, x, timesteps): + emb = self.map_noise(timesteps) + emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos + emb = self.time_embed(emb) + + x = self.proj(x) + emb + return self.mlp(x) + +class UniModMLP(nn.Module): + """ + Input: + x_num: [bs, d_numerical] + x_cat: [bs, len(categories)] + Output: + x_num_pred: [bs, d_numerical], the predicted mean for numerical data + x_cat_pred: [bs, sum(categories)], the predicted UNORMALIZED logits for categorical data + """ + def __init__( + self, d_numerical, categories, num_layers, d_token, + n_head = 1, factor = 4, bias = True, dim_t=512, use_mlp=True, **kwargs + ): + super().__init__() + self.d_numerical = d_numerical + self.categories = categories + + self.tokenizer = Tokenizer(d_numerical, categories, d_token, bias = bias) + self.encoder = Transformer(num_layers, d_token, n_head, d_token, factor) + d_in = d_token * (d_numerical + len(categories)) + self.mlp = MLPDiffusion(d_in, dim_t=dim_t, use_mlp=use_mlp) + self.decoder = Transformer(num_layers, d_token, n_head, d_token, factor) + self.detokenizer = Reconstructor(d_numerical, categories, d_token) + + self.model = nn.ModuleList([self.tokenizer, self.encoder, self.mlp, self.decoder, self.detokenizer]) + + def forward(self, x_num, x_cat, timesteps): + e = self.tokenizer(x_num, x_cat) + decoder_input = e[:, 1:, :] # ignore the first CLS token. + y = self.encoder(decoder_input) + pred_y = self.mlp(y.reshape(y.shape[0], -1), timesteps) + pred_e = self.decoder(pred_y.reshape(*y.shape)) + x_num_pred, x_cat_pred = self.detokenizer(pred_e) + x_cat_pred = torch.cat(x_cat_pred, dim=-1) if len(x_cat_pred)>0 else torch.zeros_like(x_cat).to(x_num_pred.dtype) + + return x_num_pred, x_cat_pred + + +class Precond(nn.Module): + def __init__(self, + denoise_fn, + sigma_data = 0.5, # Expected standard deviation of the training data. + net_conditioning = "sigma", + ): + super().__init__() + self.sigma_data = sigma_data + self.net_conditioning = net_conditioning + self.denoise_fn_F = denoise_fn + + def forward(self, x_num, x_cat, t, sigma): + + x_num = x_num.to(torch.float32) + + sigma = sigma.to(torch.float32) + assert sigma.ndim == 2 + if sigma.dim() > 1: # if learnable column-wise noise schedule, sigma conditioning is set to the defaults schedule of rho=7 + sigma_cond = (0.002 ** (1/7) + t * (80 ** (1/7) - 0.002 ** (1/7))).pow(7) + else: + sigma_cond = sigma + dtype = torch.float32 + + c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt() + c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt() + c_noise = sigma_cond.log() / 4 + + x_in = c_in * x_num + if self.net_conditioning == "sigma": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, c_noise.flatten()) + elif self.net_conditioning == "t": + F_x, x_cat_pred = self.denoise_fn_F(x_in, x_cat, t) + + assert F_x.dtype == dtype + D_x = c_skip * x_num + c_out * F_x.to(torch.float32) + + return D_x, x_cat_pred + + +class Model(nn.Module): + def __init__( + self, denoise_fn, + sigma_data=0.5, + precond=False, + net_conditioning="sigma", + **kwargs + ): + super().__init__() + self.precond = precond + if precond: + self.denoise_fn_D = Precond( + denoise_fn, + sigma_data=sigma_data, + net_conditioning=net_conditioning + ) + else: + self.denoise_fn_D = denoise_fn + + def forward(self, x_num, x_cat, t, sigma=None): + if self.precond: + return self.denoise_fn_D(x_num, x_cat, t, sigma) + else: + return self.denoise_fn_D(x_num, x_cat, t) + + diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/modules/transformer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/modules/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bedffe15eaa4fec85d7f686ba8741dc5d4d7ac04 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/modules/transformer.py @@ -0,0 +1,258 @@ +import numpy as np +import torch +import torch.nn as nn +import torch.nn.init as nn_init +import torch.nn.functional as F +from torch import Tensor + +import math + +class Tokenizer(nn.Module): + + def __init__(self, d_numerical, categories, d_token, bias): + super().__init__() + if categories is None: + d_bias = d_numerical + self.category_offsets = None + self.category_embeddings = None + else: + d_bias = d_numerical + len(categories) + category_offsets = torch.tensor([0] + list(categories[:-1])).cumsum(0) + self.register_buffer('category_offsets', category_offsets) + self.cat_weight = nn.Parameter(Tensor(sum(categories), d_token)) + nn.init.kaiming_uniform_(self.cat_weight, a=math.sqrt(5)) + + # take [CLS] token into account + self.weight = nn.Parameter(Tensor(d_numerical + 1, d_token)) + self.bias = nn.Parameter(Tensor(d_bias, d_token)) if bias else None + # The initialization is inspired by nn.Linear + nn_init.kaiming_uniform_(self.weight, a=math.sqrt(5)) + if self.bias is not None: + nn_init.kaiming_uniform_(self.bias, a=math.sqrt(5)) + + @property + def n_tokens(self): + return len(self.weight) + ( + 0 if self.category_offsets is None else len(self.category_offsets) + ) + + def forward(self, x_num, x_cat): + x_some = x_num if x_cat is None else x_cat + assert x_some is not None + x_num = torch.cat( + [torch.ones(len(x_some), 1, device=x_some.device)] # [CLS] + + ([] if x_num is None else [x_num]), + dim=1, + ) + + x = self.weight[None] * x_num[:, :, None] + + if x_cat is not None: + for start, end in zip(self.category_offsets, torch.cat([self.category_offsets[1:], torch.tensor([x_cat.shape[1]], device=x_cat.device)])): + if start < end: + x = torch.cat( + [x, x_cat[:, start:end].unsqueeze(1) @ self.cat_weight[start:end][None]], + dim=1, + ) + if self.bias is not None: + bias = torch.cat( + [ + torch.zeros(1, self.bias.shape[1], device=x.device), + self.bias, + ] + ) + x = x + bias[None] + + return x + + +class MultiheadAttention(nn.Module): + def __init__(self, d, n_heads, dropout, initialization = 'kaiming'): + + if n_heads > 1: + assert d % n_heads == 0 + assert initialization in ['xavier', 'kaiming'] + + super().__init__() + self.W_q = nn.Linear(d, d) + self.W_k = nn.Linear(d, d) + self.W_v = nn.Linear(d, d) + self.W_out = nn.Linear(d, d) if n_heads > 1 else None + self.n_heads = n_heads + self.dropout = nn.Dropout(dropout) if dropout else None + + for m in [self.W_q, self.W_k, self.W_v]: + if initialization == 'xavier' and (n_heads > 1 or m is not self.W_v): + # gain is needed since W_qkv is represented with 3 separate layers + nn_init.xavier_uniform_(m.weight, gain=1 / math.sqrt(2)) + nn_init.zeros_(m.bias) + if self.W_out is not None: + nn_init.zeros_(self.W_out.bias) + + def _reshape(self, x): + batch_size, n_tokens, d = x.shape + d_head = d // self.n_heads + return ( + x.reshape(batch_size, n_tokens, self.n_heads, d_head) + .transpose(1, 2) + .reshape(batch_size * self.n_heads, n_tokens, d_head) + ) + + def forward(self, x_q, x_kv, key_compression = None, value_compression = None): + + q, k, v = self.W_q(x_q), self.W_k(x_kv), self.W_v(x_kv) + for tensor in [q, k, v]: + assert tensor.shape[-1] % self.n_heads == 0 + if key_compression is not None: + assert value_compression is not None + k = key_compression(k.transpose(1, 2)).transpose(1, 2) + v = value_compression(v.transpose(1, 2)).transpose(1, 2) + else: + assert value_compression is None + + batch_size = len(q) + d_head_key = k.shape[-1] // self.n_heads + d_head_value = v.shape[-1] // self.n_heads + n_q_tokens = q.shape[1] + + q = self._reshape(q) + k = self._reshape(k) + + a = q @ k.transpose(1, 2) + b = math.sqrt(d_head_key) + attention = F.softmax(a/b , dim=-1) + + + if self.dropout is not None: + attention = self.dropout(attention) + x = attention @ self._reshape(v) + x = ( + x.reshape(batch_size, self.n_heads, n_q_tokens, d_head_value) + .transpose(1, 2) + .reshape(batch_size, n_q_tokens, self.n_heads * d_head_value) + ) + if self.W_out is not None: + x = self.W_out(x) + + return x + +class Transformer(nn.Module): + + def __init__( + self, + n_layers: int, + d_token: int, + n_heads: int, + d_out: int, + d_ffn_factor: int, + attention_dropout = 0.0, + ffn_dropout = 0.0, + residual_dropout = 0.0, + activation = 'relu', + prenormalization = True, + initialization = 'kaiming', + ): + super().__init__() + + def make_normalization(): + return nn.LayerNorm(d_token) + + d_hidden = int(d_token * d_ffn_factor) + self.layers = nn.ModuleList([]) + for layer_idx in range(n_layers): + layer = nn.ModuleDict( + { + 'attention': MultiheadAttention( + d_token, n_heads, attention_dropout, initialization + ), + 'linear0': nn.Linear( + d_token, d_hidden + ), + 'linear1': nn.Linear(d_hidden, d_token), + 'norm1': make_normalization(), + } + ) + if not prenormalization or layer_idx: + layer['norm0'] = make_normalization() + + self.layers.append(layer) + + self.activation = nn.ReLU() + self.last_activation = nn.ReLU() + # self.activation = lib.get_activation_fn(activation) + # self.last_activation = lib.get_nonglu_activation_fn(activation) + self.prenormalization = prenormalization + self.last_normalization = make_normalization() if prenormalization else None + self.ffn_dropout = ffn_dropout + self.residual_dropout = residual_dropout + self.head = nn.Linear(d_token, d_out) + + + def _start_residual(self, x, layer, norm_idx): + x_residual = x + if self.prenormalization: + norm_key = f'norm{norm_idx}' + if norm_key in layer: + x_residual = layer[norm_key](x_residual) + return x_residual + + def _end_residual(self, x, x_residual, layer, norm_idx): + if self.residual_dropout: + x_residual = F.dropout(x_residual, self.residual_dropout, self.training) + x = x + x_residual + if not self.prenormalization: + x = layer[f'norm{norm_idx}'](x) + return x + + def forward(self, x): + for layer_idx, layer in enumerate(self.layers): + is_last_layer = layer_idx + 1 == len(self.layers) + + x_residual = self._start_residual(x, layer, 0) + x_residual = layer['attention']( + # for the last attention, it is enough to process only [CLS] + x_residual, + x_residual, + ) + + x = self._end_residual(x, x_residual, layer, 0) + + x_residual = self._start_residual(x, layer, 1) + x_residual = layer['linear0'](x_residual) + x_residual = self.activation(x_residual) + if self.ffn_dropout: + x_residual = F.dropout(x_residual, self.ffn_dropout, self.training) + x_residual = layer['linear1'](x_residual) + x = self._end_residual(x, x_residual, layer, 1) + return x + + +class Reconstructor(nn.Module): + def __init__(self, d_numerical, categories, d_token): + super(Reconstructor, self).__init__() + + self.d_numerical = d_numerical + self.categories = categories + self.d_token = d_token + + self.weight = nn.Parameter(Tensor(d_numerical, d_token)) + nn.init.xavier_uniform_(self.weight, gain=1 / math.sqrt(2)) + self.cat_recons = nn.ModuleList() + + for d in categories: + recon = nn.Linear(d_token, d) + nn.init.xavier_uniform_(recon.weight, gain=1 / math.sqrt(2)) + self.cat_recons.append(recon) + + def forward(self, h): + h_num = h[:, :self.d_numerical] + h_cat = h[:, self.d_numerical:] + + recon_x_num = torch.mul(h_num, self.weight.unsqueeze(0)).sum(-1) + recon_x_cat = [] + + for i, recon in enumerate(self.cat_recons): + + recon_x_cat.append(recon(h_cat[:, i])) + + return recon_x_num, recon_x_cat diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/trainer.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..35ee6d7afff62017dcaa63ca46c3bb7b2cc246f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/tabdiff/trainer.py @@ -0,0 +1,657 @@ +import os +import glob +import time +import torch +from torch.optim.lr_scheduler import ReduceLROnPlateau +import numpy as np +import pandas as pd +import json + +from copy import deepcopy + +from utils_train import update_ema + +from tqdm import tqdm + +BAR = "==============" +def print_with_bar(log_msg): + log_msg = BAR + log_msg + BAR + if "End" in log_msg: + log_msg += "\n" + print(log_msg) + +class Trainer: + def __init__( + self, diffusion, train_iter, dataset, test_dataset, metrics, logger, + lr, weight_decay, + steps, batch_size, check_val_every, + sample_batch_size, model_save_path, result_save_path, + num_samples_to_generate=None, + lr_scheduler='reduce_lr_on_plateau', + reduce_lr_patience=100, factor=0.9, + ema_decay=0.997, + closs_weight_schedule = "fixed", + c_lambda = 1.0, + d_lambda = 1.0, + device=torch.device('cuda:1'), + ckpt_path = None, + y_only=False, + **kwargs + ): + self.y_only = y_only + self.diffusion = diffusion + self.ema_model = deepcopy(self.diffusion._denoise_fn) + for param in self.ema_model.parameters(): + param.detach_() + self.ema_num_schedule = deepcopy(self.diffusion.num_schedule) + for param in self.ema_num_schedule.parameters(): + param.detach_() + self.ema_cat_schedule = deepcopy(self.diffusion.cat_schedule) + for param in self.ema_cat_schedule.parameters(): + param.detach_() + + self.train_iter = train_iter + self.dataset = dataset + self.test_dataset = test_dataset + self.steps = steps + self.init_lr = lr + self.optimizer = torch.optim.AdamW(self.diffusion.parameters(), lr=lr, weight_decay=weight_decay) + self.ema_decay = ema_decay + self.lr_scheduler = lr_scheduler + # PyTorch >= 2.8: ReduceLROnPlateau no longer accepts `verbose=` + self.scheduler = ReduceLROnPlateau( + self.optimizer, mode="min", factor=factor, patience=reduce_lr_patience + ) + self.closs_weight_schedule = closs_weight_schedule + self.c_lambda = c_lambda + self.d_lambda = d_lambda + + self.batch_size = batch_size + self.sample_batch_size = sample_batch_size + self.num_samples_to_generate = num_samples_to_generate + self.metrics = metrics + self.logger = logger + self.check_val_every = check_val_every + + self.device = device + self.model_save_path = model_save_path + self.result_save_path = result_save_path + self.ckpt_path = ckpt_path + if self.ckpt_path is not None: + state_dicts = torch.load(self.ckpt_path, map_location=self.device) + self.diffusion._denoise_fn.load_state_dict(state_dicts['denoise_fn']) + self.diffusion.num_schedule.load_state_dict(state_dicts['num_schedule']) + self.diffusion.cat_schedule.load_state_dict(state_dicts['cat_schedule']) + print(f"Weights are loaded from {self.ckpt_path}") + + self.curr_epoch = int(os.path.basename(self.ckpt_path).split('_')[-1].split('.')[0]) if self.ckpt_path is not None else 0 + + def _anneal_lr(self, step): + frac_done = step / self.steps + lr = self.init_lr * (1 - frac_done) + for param_group in self.optimizer.param_groups: + param_group["lr"] = lr + + def _run_step(self, x, closs_weight, dloss_weight): + x = x.to(self.device) + + self.diffusion.train() + + self.optimizer.zero_grad() + + dloss, closs = self.diffusion.mixed_loss(x) + + loss = dloss_weight * dloss + closs_weight * closs + loss.backward() + self.optimizer.step() + + return dloss, closs + + def compute_loss(self): # eval loss is not weighted + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + data_iter = self.train_iter + for batch in data_iter: + x = batch.float().to(self.device) + self.diffusion.eval() + with torch.no_grad(): + batch_dloss, batch_closs = self.diffusion.mixed_loss(x) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + return mloss, gloss + + def run_loop(self): + patience = 0 + closs_weight, dloss_weight = self.c_lambda, self.d_lambda + best_loss = np.inf + best_ema_loss = np.inf + best_val_loss = np.inf + start_time = time.time() + print_with_bar(f"Starting Trainin Loop, total number of epoch = {self.steps}") + # Set up wandb's step metric + self.logger.define_metric("epoch") + self.logger.define_metric("*", step_metric="epoch") + + start_epoch = self.curr_epoch + if start_epoch > 0: + print_with_bar(f"Resuming training from epoch {start_epoch}, with validation check every {self.check_val_every} epoches") + for epoch in range (start_epoch, self.steps): + self.curr_epoch = epoch+1 + # Set up pbar + pbar = tqdm(self.train_iter, total=len(self.train_iter)) + pbar.set_description(f"Epoch {epoch+1}/{self.steps}") + + # Compute the loss weights + if self.closs_weight_schedule == "fixed": + pass + elif self.closs_weight_schedule == "anneal": + frac_done = epoch / self.steps + closs_weight = self.c_lambda * (1 - frac_done) + else: + raise NotImplementedError(f"The continuous loss weight schedule {self.closs_weight_schedule} is not implemneted") + + # Training Step + curr_dloss = 0.0 + curr_closs = 0.0 + curr_count = 0 + curr_lr = self.optimizer.param_groups[0]['lr'] + for batch in pbar: + x = batch.float().to(self.device) + batch_dloss, batch_closs = self._run_step(x, closs_weight, dloss_weight) + curr_dloss += batch_dloss.item() * len(x) + curr_closs += batch_closs.item() * len(x) + curr_count += len(x) + pbar.set_postfix({ + "lr": curr_lr, + "DLoss": np.around(curr_dloss/curr_count, 4), + "CLoss": np.around(curr_closs/curr_count, 4), + "TotalLoss": np.around((curr_dloss + curr_closs)/curr_count, 4), + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + }) + + # Log training Loss + log_dict = {} + mloss = np.around(curr_dloss / curr_count, 4) + gloss = np.around(curr_closs / curr_count, 4) + total_loss = mloss + gloss + if np.isnan(gloss): + print('Finding Nan in gaussian loss') + break + loss_dict = { + "epoch": epoch + 1, + "lr": curr_lr, + "closs_weight": closs_weight, + "dloss_weight": dloss_weight, + "loss/c_loss": gloss, + "loss/d_loss": mloss, + "loss/total_loss": total_loss + } + log_dict.update(loss_dict) + + # Log the learned noise schedules for numerical dimensions + if self.dataset.d_numerical > 0: # numerical data is not empty + num_noise_dict = {} + if self.diffusion.num_schedule.rho().dim() == 0: # non-learnable num schedule + num_noise_dict = {"num_noise/rho": self.diffusion.num_schedule.rho().item()} + else: + num_noise_dict = {f"num_noise/rho_col_{i}": value.item() for i, value in enumerate(self.diffusion.num_schedule.rho())} + log_dict.update(num_noise_dict) + + # Log the learned noise schedules for categlrical dimensions + if len(self.dataset.categories) > 0: # categorical data is not empty + cat_noise_dict = {} + if self.diffusion.cat_schedule.k().dim() == 0: # non-learnable cat schedule + cat_noise_dict = {"cat_noise/k": self.diffusion.cat_schedule.k().item()} + else: + cat_noise_dict = {f"cat_noise/k_col_{i}": value.item() for i, value in enumerate(self.diffusion.cat_schedule.k())} + log_dict.update(cat_noise_dict) + + # Adjust learning rate + if self.lr_scheduler == 'reduce_lr_on_plateau': + self.scheduler.step(total_loss) + elif self.lr_scheduler == 'anneal': + self._anneal_lr(epoch) + elif self.lr_scheduler == 'fixed': + pass + else: + raise NotImplementedError(f"LR scheduler with name '{self.lr_scheduler}' is not implemented") + + # Update EMA models + update_ema(self.ema_model.parameters(), self.diffusion._denoise_fn.parameters(), rate=self.ema_decay) + update_ema(self.ema_num_schedule.parameters(), self.diffusion.num_schedule.parameters(), rate=self.ema_decay) + update_ema(self.ema_cat_schedule.parameters(), self.diffusion.cat_schedule.parameters(), rate=self.ema_decay) + + # Save ckpt base on the best training loss + if total_loss < best_loss and self.curr_epoch > 4000: + best_loss = total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_model_{np.round(total_loss,4)}_{epoch+1}.pt')) + patience = 0 + else: + patience += 1 # increment patience if best loss is not surpassed + + # Compute and log EMA model loss + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + ema_mloss, ema_gloss = self.compute_loss() + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + ema_total_loss = ema_mloss + ema_gloss + ema_loss_dict = { + "ema_loss/c_loss": ema_gloss, + "ema_loss/d_loss": ema_mloss, + "ema_loss/total_loss": ema_total_loss + } + + # Save the best ema ckpt + if ema_total_loss < best_ema_loss and self.curr_epoch > 4000: + best_ema_loss = ema_total_loss + to_remove = glob.glob(os.path.join(self.model_save_path, f"best_ema_model_*")) + if to_remove: + os.remove(to_remove[0]) + state_dicts = { + 'denoise_fn': self.ema_model.state_dict(), + 'num_schedule':self.ema_num_schedule.state_dict(), + 'cat_schedule': self.ema_cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'best_ema_model_{np.round(ema_total_loss,4)}_{epoch+1}.pt')) + + # Evaluate Sample Quality + if (epoch+1) % self.check_val_every == 0: + state_dicts = { + 'denoise_fn': self.diffusion._denoise_fn.state_dict(), + 'num_schedule':self.diffusion.num_schedule.state_dict(), + 'cat_schedule': self.diffusion.cat_schedule.state_dict(), + } + torch.save(state_dicts, os.path.join(self.model_save_path, f'model_{epoch+1}.pt')) + + print_with_bar(f"Routine Generation Evaluation every {self.check_val_every}, currently at epoch #{epoch+1}, wiht total_loss={total_loss}.") + # 适配器训练：容器内无 Chrome，跳过 Kaleido 密度图 + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + log_dict.update(out_metrics) + print(f"Eval Resutls of the Non-EMA model:\n {out_metrics}") + + # Evaluate the EMA model + torch.save(self.ema_model.state_dict(), os.path.join(self.model_save_path, f'ema_model_{epoch+1}.pt')) + ema_out_metrics, _, _ = self.evaluate_generation(ema=True, save_metric_details=True, plot_density=_plot_density) + log_dict.update({ + "ema": ema_out_metrics, + }) + print(f"Eval Resutls of the EMA model:\n {ema_out_metrics}") + + # Submit logs + self.logger.log(log_dict) + + end_time = time.time() + print_with_bar(f"Ending Trainnig Loop, totoal training time = {end_time - start_time}") + self.logger.log({ + 'training_time': end_time - start_time + }) + + def report_test(self, num_runs): + save_dir = self.result_save_path + + shape_ = [] + trend_ = [] + mle_ = [] + c2st_ = [] + for i in range(num_runs): + print_with_bar(f"GENERAL Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + shape_.append(out_metrics["density/Shape"]) + trend_.append(out_metrics["density/Trend"]) + mle_.append(out_metrics["mle"]) + c2st_.append(out_metrics["c2st"]) + # Save samples for quality evaluation + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + shape_ = np.array(shape_) + trend_ = np.array(trend_) + mle_ = np.array(mle_) + c2st_ = np.array(c2st_) + + shape_error = (1 - shape_)*100 + trend_error = (1 - trend_)*100 + c2st_percent = c2st_ * 100 + + all_results = pd.DataFrame({ + "shape": shape_error, + "trend": trend_error, + "mle": mle_, + "c2st": c2st_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "shape", + "trend", + "mle", + "c2st", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def report_test_dcr(self, num_runs): + save_dir = self.result_save_path + + dcr_ = [] + dcr_real_ = [] + dcr_test_ = [] + for i in range(num_runs): + print_with_bar(f"DCR Evaluation Run {i}") + out_metrics, extras, syn_df = self.evaluate_generation() + print(f"Results of Run {i} are: \n{out_metrics}") + dcr_.append(out_metrics["dcr"]) + dcr_real_.append(extras["dcr_real"]) + dcr_test_.append(extras["dcr_test"]) + save_path = os.path.join(save_dir, "all_samples") + if not os.path.exists(save_path): + os.makedirs(save_path) + syn_df.to_csv(os.path.join(save_path, f"samples_{i}.csv"), index=False) + + dcr_ = np.array(dcr_) + + dcr_percent = dcr_ * 100 + + all_results = pd.DataFrame({ + "dcr": dcr_percent, + }) + avg = all_results.mean(axis=0).round(3) + std = all_results.std(axis=0).round(3) + avg_std = pd.concat([avg, std], axis=1, ignore_index=True) + avg_std.columns = ["avg", "std"] + avg_std.index = [ + "dcr", + ] + + # Savings + all_results.to_csv(f"{save_dir}/all_results.csv", index=True) + avg_std.to_csv(f"{save_dir}/avg_std.csv", index=True) + dcr_real = np.concatenate(dcr_real_, axis=0) + dcr_test = np.concatenate(dcr_test_, axis=0) + dcr_df = pd.DataFrame({ + "dcr_real": dcr_real, + "dcr_test": dcr_test + }) + dcr_df.to_csv(f"{save_dir}/dcr.csv", index=False) + + print_with_bar(f"The AVG over {num_runs} runs are: \n{avg_std}") + + def test(self): + _plot_density = os.environ.get("TABDIFF_ADAPTER_TRAIN", "").strip().lower() not in ("1", "true", "yes") + out_metrics, _, _ = self.evaluate_generation(save_metric_details=True, plot_density=_plot_density) + print_with_bar(f"Results of the test are: \n{out_metrics}") + self.logger.log(out_metrics) + print(out_metrics) + + def evaluate_generation(self, save_metric_details=False, plot_density=False, ema=False): + self.diffusion.eval() + + # Sample a synthetic table + num_samples = self.num_samples_to_generate if self.num_samples_to_generate else self.metrics.real_data_size # By default, num_samples_to_generate is not specified. In this case, we generate the same number of samples as the real dataset. This approach is consistently used across all experiments in the paper. + syn_df = self.sample_synthetic(num_samples, ema=ema) + + # Save the sample + save_path = os.path.join(self.result_save_path, str(self.curr_epoch), "ema" if ema else "") + if not os.path.exists(save_path): + os.makedirs(save_path) + path = os.path.join(save_path, "samples.csv") + syn_df.to_csv(path, index=False) + print( + f"Samples are saved at {path}" + ) + + # 流水线仅需要 CSV：跳过 MLE/C2ST 等（合成表为字符串类别时 MLE 会报错） + if os.environ.get("TABDIFF_ADAPTER_SAMPLE_ONLY", "").strip().lower() in ("1", "true", "yes"): + return {}, {}, syn_df + + # Compute evaluation metrics on the sample + syn_df_loaded = pd.read_csv(os.path.join(save_path, "samples.csv")) # In the original tabsyn code, syn_data is implicitly casted into float.64 when it gets loaded with pd.read_csv in the evaluation script. If we don't cast, the density evluation for some columns (especially those with tailed and peaked distribution) will collapse. + out_metrics, extras = self.metrics.evaluate(syn_df_loaded) + + # Save metrics and metric details + path = os.path.join(save_path, "all_results.json") + with open(path, "w") as json_file: + json.dump(out_metrics, json_file, indent=4, separators=(", ", ": ")) # always locally save the output metrics + if save_metric_details: + for name, extra in extras.items(): + if isinstance(extra, pd.DataFrame): + extra.to_csv(os.path.join(save_path, f"{name}.csv")) + elif isinstance(extra, dict): + with open(os.path.join(save_path, f"{name}.json"), "w") as json_file: + json.dump(extra, json_file, indent=4, separators=(", ", ": ")) + else: + raise NotImplementedError(f"Extra file generated during evaluations has type {type(extra)}, and code to save this type of file is not implemented") + + # Plot density figures + if plot_density: + img = self.metrics.plot_density(syn_df_loaded) + path = os.path.join(save_path, "density_plots.png") + img.save(path) + print( + f"The density plots are saved at {path}" + ) + return out_metrics, extras, syn_df + + + def sample_synthetic(self, num_samples, keep_nan_samples=True, ema=False): + if ema: + curr_model, curr_num_schedule, curr_cat_schedule = self.to_ema_model() + info = self.metrics.info + + print_with_bar(f"Starting Sampling, total samples to generate = {num_samples}") + start_time = time.time() + + syn_data = self.diffusion.sample_all(num_samples, self.sample_batch_size, keep_nan_samples=keep_nan_samples) + print(f"Shape of the generated sample = {syn_data.shape}") + + if keep_nan_samples: + num_all_zero_row = (syn_data.sum(dim=1) == 0).sum() + if num_all_zero_row: + print(f"The generated samples contain {num_all_zero_row} Nan instances!!!") + self.logger.log({ + 'num_Nan_sample': num_all_zero_row + }) + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if self.y_only: + if info['task_type'] == 'binclass': + syn_data = cat_inverse(syn_data) + else: + syn_data = num_inverse(syn_data) + syn_df = pd.DataFrame() + syn_df[info['column_names'][info['target_col_idx'][0]]] = syn_data[:, 0] + else: + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + end_time = time.time() + print_with_bar(f"Ending Sampling, totoal sampling time = {end_time - start_time}") + + if ema: + self.to_model(curr_model, curr_num_schedule, curr_cat_schedule) + + return syn_df + + def to_ema_model(self): + curr_model = self.diffusion._denoise_fn + curr_num_schedule = self.diffusion.num_schedule + curr_cat_schedule = self.diffusion.cat_schedule + self.diffusion._denoise_fn = self.ema_model # temporarily install the ema parameters into the model + self.diffusion.num_schedule = self.ema_num_schedule + self.diffusion.cat_schedule = self.ema_cat_schedule + + return curr_model, curr_num_schedule, curr_cat_schedule + + def to_model(self, curr_model, curr_num_schedule, curr_cat_schedule): + self.diffusion._denoise_fn = curr_model # give back the parameters + self.diffusion.num_schedule = curr_num_schedule + self.diffusion.cat_schedule = curr_cat_schedule + + def test_impute(self, trail_start, trial_size, resample_rounds, impute_condition, imputed_sample_save_dir, w_num, w_cat): + self.diffusion.eval() + + info = self.metrics.info + task_type = info['task_type'] + d_numerical, categories = self.dataset.d_numerical, self.dataset.categories + num_mask_idx, cat_mask_idx = [], [] + X_train = self.dataset.X + X_train = X_train + x_num_train, x_cat_train = X_train[:,:d_numerical], X_train[:,d_numerical:] + + if task_type == 'binclass': # for cat cols, push the masked col to [MASK] + cat_mask_idx += [0] + else: # for num cols, set the masked col to the col mean + num_mask_idx += [0] + avg = x_num_train[:, num_mask_idx].mean(0).to(self.device) + + with torch.no_grad(): + + for trial in range(trail_start, trail_start+trial_size): + print_with_bar(f"Imputing trial {trial}") + X_test = self.test_dataset.X + X_test = deepcopy(X_test).to(self.device) + x_num_test, x_cat_test = X_test[:, :d_numerical], X_test[:, d_numerical:].long() + + # Apply mask to x_0 + if num_mask_idx: + x_num_test[:, num_mask_idx] = avg + if cat_mask_idx: + x_cat_test[:, cat_mask_idx] = torch.tensor(categories, dtype=x_cat_test.dtype, device=x_cat_test.device)[cat_mask_idx] + + # Sample imputed tables + syn_data = self.diffusion.sample_impute(x_num_test, x_cat_test, num_mask_idx, cat_mask_idx, resample_rounds, impute_condition, w_num, w_cat) + print(f"Shape of the imputed sample = {syn_data.shape}") + + # Recover tables + num_inverse = self.dataset.num_inverse + int_inverse = self.dataset.int_inverse + cat_inverse = self.dataset.cat_inverse + + if torch.any((syn_data[:, d_numerical+1:]).max(dim=0).values > (x_cat_train[:,1:]).max(dim=0).values): # if the test set contains categories not presented in the train set, we can not do cat_inverse. So we implement a patch that set those columns to the same as the train set + print("Test set contains extra categories, and so does imputed syn data. We cannot do cat_inverse. So we set the cat columns as the same as the train set") + syn_data[:, d_numerical+1:] = x_cat_train[:syn_data.shape[0],1:] + + + syn_num, syn_cat, syn_target = split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse) + syn_df = recover_data(syn_num, syn_cat, syn_target, info) + + idx_name_mapping = info['idx_name_mapping'] + idx_name_mapping = {int(key): value for key, value in idx_name_mapping.items()} + + syn_df.rename(columns = idx_name_mapping, inplace=True) + + # Save imputed samples + os.makedirs(imputed_sample_save_dir) if not os.path.exists(imputed_sample_save_dir) else None + print(f"Imputed samples are saved to {imputed_sample_save_dir}/{trial}.csv") + syn_df.to_csv(f'{imputed_sample_save_dir}/{trial}.csv', index = False) + +def _as_numpy_float32(x): + """Inverse 变换可能返回 Tensor；统一为 numpy float32（含 0 列）。""" + if x is None: + return np.array([], dtype=np.float32) + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + return np.asarray(x, dtype=np.float32) + + +@torch.no_grad() +def split_num_cat_target(syn_data, info, num_inverse, int_inverse, cat_inverse): + task_type = info['task_type'] + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + n_num_feat = len(num_col_idx) + n_cat_feat = len(cat_col_idx) + + if task_type == 'regression': + n_num_feat += len(target_col_idx) + else: + n_cat_feat += len(target_col_idx) + + syn_num = syn_data[:, :n_num_feat] + syn_cat = syn_data[:, n_num_feat:] + + if n_num_feat > 0: + syn_num = _as_numpy_float32(num_inverse(syn_num)) + syn_num = _as_numpy_float32(int_inverse(syn_num)) + else: + syn_num = np.zeros((syn_data.shape[0], 0), dtype=np.float32) + syn_cat = cat_inverse(syn_cat) + + + if info['task_type'] == 'regression': + syn_target = syn_num[:, :len(target_col_idx)] + syn_num = syn_num[:, len(target_col_idx):] + + else: + print(syn_cat.shape) + syn_target = syn_cat[:, :len(target_col_idx)] + syn_cat = syn_cat[:, len(target_col_idx):] + + return syn_num, syn_cat, syn_target + +def recover_data(syn_num, syn_cat, syn_target, info): + + num_col_idx = info['num_col_idx'] + cat_col_idx = info['cat_col_idx'] + target_col_idx = info['target_col_idx'] + + + idx_mapping = info['idx_mapping'] + idx_mapping = {int(key): value for key, value in idx_mapping.items()} + + syn_df = pd.DataFrame() + + if info['task_type'] == 'regression': + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + + else: + for i in range(len(num_col_idx) + len(cat_col_idx) + len(target_col_idx)): + if i in set(num_col_idx): + syn_df[i] = syn_num[:, idx_mapping[i]] + elif i in set(cat_col_idx): + syn_df[i] = syn_cat[:, idx_mapping[i] - len(num_col_idx)] + else: + syn_df[i] = syn_target[:, idx_mapping[i] - len(num_col_idx) - len(cat_col_idx)] + + return syn_df \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/utils_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2c84a1d209441b244c84e8f3c3e08203d496960e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_runtime/utils_train.py @@ -0,0 +1,198 @@ +import numpy as np +import os + +import src +from torch.utils.data import Dataset + +import torch + + +class TabularDataset(Dataset): + def __init__(self, X_num, X_cat): + self.X_num = X_num + self.X_cat = X_cat + + def __getitem__(self, index): + this_num = self.X_num[index] + this_cat = self.X_cat[index] + + sample = (this_num, this_cat) + + return sample + + def __len__(self): + return self.X_num.shape[0] + +class TabDiffDataset(Dataset): + def __init__(self, dataname, data_dir, info, isTrain=True, y_only=False, dequant_dist='none', int_dequant_factor=0.0): + self.dataname = dataname + self.data_dir = data_dir + self.info = info + self.isTrain = isTrain + + X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse = preprocess(data_dir, y_only, dequant_dist, int_dequant_factor, task_type = info['task_type'], inverse=True) + categories = np.array(categories) + + X_train_num, _ = X_num + X_train_cat, _ = X_cat + + X_train_num, X_test_num = X_num + X_train_cat, X_test_cat = X_cat + + X_train_num, X_test_num = torch.tensor(X_train_num).float(), torch.tensor(X_test_num).float() + X_train_cat, X_test_cat = torch.tensor(X_train_cat), torch.tensor(X_test_cat) + + self.X = torch.cat((X_train_num, X_train_cat), dim=1) if isTrain else torch.cat((X_test_num, X_test_cat), dim=1) + self.num_inverse = num_inverse + self.int_inverse = int_inverse + self.cat_inverse = cat_inverse + self.d_numerical = d_numerical + self.categories = categories + + def __getitem__(self, index): + return self.X[index] + + def __len__(self): + return self.X.shape[0] + +def preprocess(dataset_path, y_only=False, dequant_dist='none', int_dequant_factor=0.0, task_type = 'binclass', inverse = False, cat_encoding = None, concat = True): + + T_dict = {} + + T_dict['normalization'] = "quantile" + T_dict['num_nan_policy'] = 'mean' + T_dict['cat_nan_policy'] = None + T_dict['cat_min_frequency'] = None + T_dict['cat_encoding'] = cat_encoding + T_dict['y_policy'] = "default" + T_dict['dequant_dist'] = dequant_dist + T_dict['int_dequant_factor'] = int_dequant_factor + + T = src.Transformations(**T_dict) + + dataset = make_dataset( + data_path = dataset_path, + T = T, + task_type = task_type, + change_val = False, + concat = concat, + y_only = y_only, + ) + + if cat_encoding is None: + X_num = dataset.X_num + X_cat = dataset.X_cat + + X_train_num, X_test_num = X_num['train'], X_num['test'] + if X_cat is None: + X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64) + X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64) + categories = [] + else: + X_train_cat, X_test_cat = X_cat['train'], X_cat['test'] + categories = src.get_categories(X_train_cat) + d_numerical = X_train_num.shape[1] + + X_num = (X_train_num, X_test_num) + X_cat = (X_train_cat, X_test_cat) + + + if inverse: + num_inverse = dataset.num_transform.inverse_transform if dataset.num_transform is not None else lambda x: x + int_inverse = dataset.int_transform.inverse_transform if dataset.int_transform is not None else lambda x: x + cat_inverse = dataset.cat_transform.inverse_transform if dataset.cat_transform is not None else lambda x: x + + return X_num, X_cat, categories, d_numerical, num_inverse, int_inverse, cat_inverse + else: + return X_num, X_cat, categories, d_numerical + else: + return dataset + + +def update_ema(target_params, source_params, rate=0.999): + """ + Update target parameters to be closer to those of source parameters using + an exponential moving average. + :param target_params: the target parameter sequence. + :param source_params: the source parameter sequence. + :param rate: the EMA rate (closer to 1 means slower). + """ + for target, source in zip(target_params, source_params): + target.detach().mul_(rate).add_(source.detach(), alpha=1 - rate) + + + +def concat_y_to_X(X, y): + if X is None: + return y.reshape(-1, 1) + return np.concatenate([y.reshape(-1, 1), X], axis=1) + + +def make_dataset( + data_path: str, + T: src.Transformations, + task_type, + change_val: bool, + concat = True, + y_only = False, +): + + # classification + if task_type == 'binclass' or task_type == 'multiclass': + has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) + X_cat = {} if (has_cat or concat) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + X_num[split] = X_num_t + if X_cat is not None: + if concat: + X_cat_t = concat_y_to_X(X_cat_t, y_t) + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + else: + # regression + X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None + X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None + y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None + + for split in ['train', 'test']: + X_num_t, X_cat_t, y_t = src.read_pure_data(data_path, split) + if y_only: + X_num_t = X_num_t[:, :0] + X_cat_t = X_cat_t[:, :0] + if X_num is not None: + if concat: + X_num_t = concat_y_to_X(X_num_t, y_t) + X_num[split] = X_num_t + if X_cat is not None: + X_cat[split] = X_cat_t + if y is not None: + y[split] = y_t + + info = src.load_json(os.path.join(data_path, 'info.json')) + int_col_idx_wrt_num = info['int_col_idx_wrt_num'] + + if y_only: + int_col_idx_wrt_num = [] + D = src.Dataset( + X_num, + X_cat, + y, + int_col_idx_wrt_num, + y_info={}, + task_type=src.TaskType(info['task_type']), + n_classes=info.get('n_classes') + ) + + if change_val: + D = src.change_val(D) + + return src.transform_dataset(D, T, None) \ No newline at end of file diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_train.py b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_train.py new file mode 100644 index 0000000000000000000000000000000000000000..478d9743661f9c71e455efb27fb0c81c770815c3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/_tabdiff_train.py @@ -0,0 +1,78 @@ + +import os, shutil, subprocess, sys +td = r"/workspace/TabDiff" +name = r"pipeline_n11" +src = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11" +rt = r"/work/output-Benchmark-trainonly-v1/n11/tabdiff/tabdiff-n11-20260521_053348/_tabdiff_runtime" +shutil.rmtree(rt, ignore_errors=True) + +def _ignore(_, names): + skip = {"__pycache__", "data", "synthetic", "result", "results", "ckpt"} + return [n for n in names if n in skip or n.endswith(".pyc")] + +shutil.copytree(td, rt, ignore=_ignore) + +def _replace_once(path, old, new): + text = open(path, "r", encoding="utf-8").read() + if old not in text: + raise RuntimeError(f"patch anchor not found in {path}") + text = text.replace(old, new, 1) + with open(path, "w", encoding="utf-8") as f: + f.write(text) + +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_train_num, X_test_num = X_num['train'], X_num['test']\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n \n categories = src.get_categories(X_train_cat)\n", + " X_train_num, X_test_num = X_num['train'], X_num['test']\n if X_cat is None:\n X_train_cat = np.empty((X_train_num.shape[0], 0), dtype=np.int64)\n X_test_cat = np.empty((X_test_num.shape[0], 0), dtype=np.int64)\n categories = []\n else:\n X_train_cat, X_test_cat = X_cat['train'], X_cat['test']\n categories = src.get_categories(X_train_cat)\n", +) +_replace_once( + os.path.join(rt, "utils_train.py"), + " X_cat = {} if os.path.exists(os.path.join(data_path, 'X_cat_train.npy')) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", + " has_cat = os.path.exists(os.path.join(data_path, 'X_cat_train.npy'))\n X_cat = {} if (has_cat or concat) else None\n X_num = {} if os.path.exists(os.path.join(data_path, 'X_num_train.npy')) else None\n y = {} if os.path.exists(os.path.join(data_path, 'y_train.npy')) else None\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " num_workers=1,\n", + " num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +_replace_once( + os.path.join(rt, "src", "data.py"), + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=1)\n", + " loader = torch.utils.data.DataLoader(torch_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=int(os.environ.get('TABDIFF_NUM_WORKERS', '0')))\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n ## Load training data\n", + " if os.environ.get(\"TABDIFF_ADAPTER_TRAIN\", \"\").strip() and args.mode == \"train\":\n raw_config[\"train\"][\"main\"][\"check_val_every\"] = int(raw_config[\"train\"][\"main\"][\"steps\"])\n\n _train_batch = os.environ.get(\"TABDIFF_BATCH_SIZE\", \"\").strip() or os.environ.get(\"TABDIFF_TRAIN_BATCH_SIZE\", \"\").strip()\n if _train_batch:\n raw_config[\"train\"][\"main\"][\"batch_size\"] = max(1, int(_train_batch))\n _sample_batch = os.environ.get(\"TABDIFF_SAMPLE_BATCH_SIZE\", \"\").strip()\n if _sample_batch:\n raw_config[\"sample\"][\"batch_size\"] = max(1, int(_sample_batch))\n _train_lr = os.environ.get(\"TABDIFF_LR\", \"\").strip() or os.environ.get(\"TABDIFF_LEARNING_RATE\", \"\").strip()\n if _train_lr:\n raw_config[\"train\"][\"main\"][\"lr\"] = float(_train_lr)\n _num_timesteps = os.environ.get(\"TABDIFF_NUM_TIMESTEPS\", \"\").strip() or os.environ.get(\"TABDIFF_TIMESTEPS\", \"\").strip()\n if _num_timesteps:\n raw_config[\"diffusion_params\"][\"num_timesteps\"] = max(1, int(_num_timesteps))\n\n ## Load training data\n", +) +_replace_once( + os.path.join(rt, "tabdiff", "main.py"), + " num_workers = 4,\n", + " num_workers = int(os.environ.get('TABDIFF_NUM_WORKERS', '0')),\n", +) +dst_data = os.path.join(rt, "data", name) +dst_syn = os.path.join(rt, "synthetic", name) +shutil.rmtree(dst_data, ignore_errors=True) +os.makedirs(os.path.dirname(dst_data), exist_ok=True) +shutil.copytree(src, dst_data) +os.makedirs(dst_syn, exist_ok=True) +for fn in ("real.csv", "test.csv", "val.csv"): + shutil.copy(os.path.join(src, fn), os.path.join(dst_syn, fn)) +os.chdir(rt) +os.environ["PYTHONPATH"] = rt + os.pathsep + os.environ.get("PYTHONPATH", "") +os.environ["TABDIFF_SMOKE_STEPS"] = "800" +os.environ["TABDIFF_STEPS"] = "800" +os.environ["TABDIFF_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_TRAIN_BATCH_SIZE"] = "1024" +os.environ["TABDIFF_SAMPLE_BATCH_SIZE"] = "512" +os.environ["TABDIFF_LR"] = "0.001" +os.environ["TABDIFF_LEARNING_RATE"] = "0.001" +os.environ["TABDIFF_NUM_TIMESTEPS"] = "100" +os.environ["TABDIFF_TIMESTEPS"] = "100" +os.environ["TABDIFF_NUM_WORKERS"] = "0" +os.environ["TABDIFF_ADAPTER_TRAIN"] = "1" +subprocess.check_call([ + sys.executable, "-m", "tabdiff.main", + "--dataname", name, "--mode", "train", "--gpu", "0", + "--no_wandb", "--exp_name", r"adapter_learnable", +]) diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/bo_combo.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..8a78306e2f8bca03bfda766a70b730a590929e74 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79a6f056d9337d674376179f0308922eae99a3fa3d028075fa63728873a0678c +size 126 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/gen_20260521_054153.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/gen_20260521_054153.log new file mode 100644 index 0000000000000000000000000000000000000000..e1b3c4b4c5b3f225c0c201cdbc18c155544b741e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/gen_20260521_054153.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:261fea9c10f8e03d81bf82fe7fc4086d9f355f7624331846fbb80a769da8dd1b +size 25601 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/input_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..f93aa1ae212ec9a9c05d8e83ded7fad3b998a73e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72149d5374a893bfa5c45ba770dafd723399f1e8e700b03f797a0cf5c65dbc4 +size 1360 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/models_tabdiff/trained.pt b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/models_tabdiff/trained.pt new file mode 100644 index 0000000000000000000000000000000000000000..bc319a55a9c7152a1137f968b6fa1c735125ad11 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/models_tabdiff/trained.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:306660b8aad4549267c68390b017e15ddb9bc06a02c80515eb72a97ae31a81eb +size 74 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..657380e29a276b82aa19aeeaf76dfa65dff1c8ff --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2946591b5413f134b5ab506e52854e44296bf65e8c66c86f891ea4ecaec63a21 +size 6099 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/run_config.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..0a8bf4835d85f8146a06f540aa854ad686f1f6e0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa320df874b4e037deeac28a51a48669178dba450ec51c4ef2949f4f8678b434 +size 2434 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/runtime_result.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..6345d5578d7e77cd59011671834391d788220926 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d29356b8a023a4e6c20c865a0b125ea4be74f354a24aeef64e1d06a2388ac500 +size 916 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/tabdiff/adapter_report.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/tabdiff/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..1e30ff6c43604adbca5c175ffba649097a890e7f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/tabdiff/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da978ba14a6f4025f9e4b7751da398087df7d25c4b08942f0b420c616a57b1fc +size 323 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/tabdiff/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/tabdiff/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/tabdiff/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/tabdiff/model_input_manifest.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/tabdiff/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2db53e308d098cdd51790891a82a27bab33d8694 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/staged/tabdiff/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72606aea675c78590398da0c9179419c335b5171ccf82f8e0c0e7412d8500b2d +size 6298 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabdiff-n11-15215-20260521_054153.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabdiff-n11-15215-20260521_054153.csv new file mode 100644 index 0000000000000000000000000000000000000000..899be2fc8007f02f909bea9dbdba927d575b6150 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabdiff-n11-15215-20260521_054153.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cab139b8d70a938fbc8cc384bd4811951824bd6c1e3bfe1b49852f7e307a65f7 +size 1557082 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabdiff_train_meta.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabdiff_train_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..c4e7110b14a039875ef5649a2a6581804ba47d7c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabdiff_train_meta.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9239295a76f1d966db4907b47735aeac1d05f743d43918eea0bbd5d097218bb +size 535 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/X_num_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..f8ab56193672162e453ef2c46d259311a5c82f4b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25f5aa50a1a26a08e947c0d7e7b819cab744a982f03e91fc4aa148b107aa8ce8 +size 76248 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/X_num_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..9dd7e540ba4068430122140bcc98f17190dbb0a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e325cf01927e570f62f1a49f4bc2822dc29b1aafe3c8be5d3d89a01d2df6e5ec +size 608728 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/X_num_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/X_num_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..21557e1a0062bdd9436fc2ef620880f1aa7991ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/X_num_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998d48896c06f1673a7af7e251054b45fab2ff2040dfe0a9206235e10f9001cb +size 76168 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/info.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/info.json new file mode 100644 index 0000000000000000000000000000000000000000..eae0ccc21e3c6fd577e4123f57ee19c528f4fe5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ca6bdc5c40407a8049688a9e9f8ff98bdbd8d26499d907751f29c09607145b +size 2199 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/real.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/staged_features.json b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/test.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/train.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/val.csv b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/y_test.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..da3aeebbed3b3846e45f08cdb1ab683db9d15ce2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:825ce3e2cdf78c6714903278473b81851367919d7f5810d9c1c71c9f02633540 +size 15352 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/y_train.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b523a92a6e5006ca3c440fc79afd02297be1f7ee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95905dd042874f4c4e7523f7525b2864e220ef3c6fcdc58877ca3944645899cf +size 121848 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/y_val.npy b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/y_val.npy new file mode 100644 index 0000000000000000000000000000000000000000..084d7578e6a5ada1f4e9d587e69c1beb323a7d2d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/tabular_bundle/pipeline_n11/y_val.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:babb398ab19fc90389945862a1227d04f9aa48293af50c7091056137a6d2f2f5 +size 15336 diff --git a/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/train_20260521_053348.log b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/train_20260521_053348.log new file mode 100644 index 0000000000000000000000000000000000000000..84b5dd005fdc6592e9e66b455d968b945cfa5651 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabdiff/runs/tabdiff-n11-20260521_053348/train_20260521_053348.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:219c3e790d8e4f65b26f0f7b0888d57931fc69d98943fecfe019be6c1e60a399 +size 2545348 diff --git a/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/final_recommendation.json b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/final_recommendation.json new file mode 100644 index 0000000000000000000000000000000000000000..79754e789ec73615f97cd6f1d8789380f4a636aa --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/final_recommendation.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e37293a5e52d084bde67c466333adb4d430466926f81e52de02c5273494f8d57 +size 2622 diff --git a/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase1_manifest.json b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase1_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..30c972346c7248efd6c443fe4a3224b34f45f924 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase1_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc932caa6b7451fbfb52a0e98ff11dead38b41d95569d945ff35bebc4d6fea4f +size 1740 diff --git a/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase1_scores.csv b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase1_scores.csv new file mode 100644 index 0000000000000000000000000000000000000000..be9baa036acb451e27e9d2ae81ddd56ed4aaa70a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase1_scores.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24db26dd5c91c0b93678b111a9931038bed133402023de1a6f76f59e2bfa0512 +size 2989 diff --git a/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase1_summary.json b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase1_summary.json new file mode 100644 index 0000000000000000000000000000000000000000..5d442cd7602aeaecfa3c690ba6fff55eca699967 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase1_summary.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b92f3bdb12cdd0c3f929ccec22b4151ffce3f4406a6e22d716f5846fbbef9133 +size 575 diff --git a/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_manifest.json b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..54e5ed18b651819d5ef4795374b530b9ced8b7e1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cd1ee95f78f4fed74da71935e9a7b6d6bd29046caae459b5b5f2f1a0f63e3d6 +size 1650 diff --git a/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_report.csv b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_report.csv new file mode 100644 index 0000000000000000000000000000000000000000..9e5ee6f79f8bbc78f87d64126350ac7cbae8fcc7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_report.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94b540139d3096dab61156040e187fcada0523f3afce0aedee26dd0d5de988ba +size 846 diff --git a/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_report.json b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_report.json new file mode 100644 index 0000000000000000000000000000000000000000..f10b81485f16999a808342f9aa117d0cb26d87a0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:912eeef28fe96cb753863f63345a00a7cac9f13a95dc43a9d8b38e67712e44ff +size 2836 diff --git a/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_scores.csv b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_scores.csv new file mode 100644 index 0000000000000000000000000000000000000000..d652822f422f5ab5469dfd240eeacee0a8a2e21b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/phase2_scores.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a68748e62e361e5164bb662969602b8ca1f4b57369af3209407be60d63aef63 +size 2523 diff --git a/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/sample_manifest.json b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/sample_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..dd594961832e37140d1d8afd66fa2643ace3f7b7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/_rf_tuning/sample_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd6f3afab8ac6c1a092fbab096c091d56f5085ba7624e4cd2441f161e6a06d2b +size 635 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..f987ceb28cfefc93355cbaf59a45d6ba3b0ad1f6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_234105" +dataname = "tabsyn_n11_tabsyn_n11_20260520_234105" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_234105/tabsyn-n11-15215-20260520_234157.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..5de70a50389fd3cff7f9cc8a52edf05e8f7a29d8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_234105" +dataname = "tabsyn_n11_tabsyn_n11_20260520_234105" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 10 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..5be83c6401118c7bf6c17b45fe05e5a6480f5cd6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2c259158b14124a7cc758f83c1c4b68f6025ea448c4851bfda0172ee9f3d7e2 +size 94 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/info.json new file mode 100644 index 0000000000000000000000000000000000000000..07aea465e309a88d43bc6e069d041f9b61ae1b88 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d89e482c7eeb33bae2cb6fcbae78a66cca684a91fab35d033984e6b74847e50a +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/data/tabsyn_n11_tabsyn_n11_20260520_234105/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/gen_20260520_234157.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/gen_20260520_234157.log new file mode 100644 index 0000000000000000000000000000000000000000..b3720034197551104a4c84427332970ee0f3ea50 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/gen_20260520_234157.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfb4d6f1bab36d3c1a6a82c6c6cfc690928f3fd5d06741f7468b9dfc0c280777 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2791b2cf321145a322b8b62cbfcf3a7afd959334 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e4eeef7c1acc96326be912c943a1a674330b25dbad89b532ea26c8a5b2aaa1d +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..0575582af94aa4cc1f2499688dcf2c6eb8f03d49 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bca094ee66013900b6cccf7cf3f690f1fe8845a8d47b2b9602a62a3c105c699b +size 2594 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..ba4978b34555d37673a57999d616fe52ba0ae9b6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74b1f3340820234cd8e2861addf808c00a027fb5fc184bd063134b9abeda6d27 +size 881 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..bdb544db49ac39f8961056bb593aaa84beb90309 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4aa02c8c14a9aa5951c4786251a48d32ea0649f89496cadbae8187e4043d659 +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b1f0cb729f78580054b607e69b4eb7adb5cdc0c2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7855f34ee2e040b4085a5342a52ddb482d506b358b84d1b6598381cae77212de +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/synthetic/tabsyn_n11_tabsyn_n11_20260520_234105/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/synthetic/tabsyn_n11_tabsyn_n11_20260520_234105/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/synthetic/tabsyn_n11_tabsyn_n11_20260520_234105/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/synthetic/tabsyn_n11_tabsyn_n11_20260520_234105/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/synthetic/tabsyn_n11_tabsyn_n11_20260520_234105/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/synthetic/tabsyn_n11_tabsyn_n11_20260520_234105/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn-n11-15215-20260520_234157.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn-n11-15215-20260520_234157.csv new file mode 100644 index 0000000000000000000000000000000000000000..781845b752cbef7f643077015aed7b2621ca4092 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn-n11-15215-20260520_234157.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0558526df7d11a167a6e69738d24d3cd817b2b725f337d6ead8b9af38e23576d +size 1584759 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234105/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234105/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..ae6d0a8a091b2335dc7babf3def9e96f56010f46 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234105/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d64a7f0aceaebf830bb7d38c6ec65644ed17080d28f070424b55c0254666303 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234105/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234105/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..80f74451ab0d7792ec136d7927ea968d2db76a0f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234105/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa84f7b7dd7c7b17923263137b9d505db30a4c1922b9cfa134f5cc61510a4573 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..06577b0fa59ac5f955978a1faa48900c240d9afb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:683a34d1274607827b28e514929fbca1b5889d279fa6a2d6a920538317171283 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..ecc9359417124b9da7819fd7bd2c5734ad467819 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0dd2ca8a964cd86b328dd0b7ed1ae213879853fbc200147ec10f4cd85d8fdf4b +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..087c6eb1da2a5d0ae25fc423b822932851f34e09 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8670db4c49e448f5d1ab621b233208ee29edc56286f4910198d783e65ef8b43 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..94e1f410feb09b7ccd8c7280379a6446f95a1e1b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234105/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32edbf974faba760477b3e2f31b3d69d6733a83cc830653dc804cb67740160bd +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/train_20260520_234106.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/train_20260520_234106.log new file mode 100644 index 0000000000000000000000000000000000000000..e245cb29ceae8709ac43377eadda5c587b2697a2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234105/train_20260520_234106.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f81e406eff46f79233dcb362e41cbcb5456633593eb25df883aafb2083466cc2 +size 37809 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..755dcaa2653cae770c44a617a177f6cb2bdbc3a1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_234313" +dataname = "tabsyn_n11_tabsyn_n11_20260520_234313" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_234313/tabsyn-n11-15215-20260520_234708.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..935e9294a86ff920b99a617138ba1cdbbee60ee3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_234313" +dataname = "tabsyn_n11_tabsyn_n11_20260520_234313" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 30 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..401494c19f6df82743f4f17019d484c49e7b6b3e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b32ec0985113c7c2d3193acd722073174570cf3882109ba5d0723fe51ad172c8 +size 92 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/info.json new file mode 100644 index 0000000000000000000000000000000000000000..935e7c5f1f178cf6e5eef60cab2e785d5a6401ca --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4fc6f333733de11a474537947a4d29b648420bde324785919dcb6eef5913fd1 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/data/tabsyn_n11_tabsyn_n11_20260520_234313/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/gen_20260520_234708.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/gen_20260520_234708.log new file mode 100644 index 0000000000000000000000000000000000000000..e3ffaa1cc6c23456fb5e268b8291ecb8fbf21fab --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/gen_20260520_234708.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d08460aaa7ac47ba0593fcb14275740da59a0dc648c7694123a3b7bd9403a829 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..80c5512ee9e94f0d09994f09a72747c74b66708e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b2c750dbf297a9a620961a8c8b5af8c59b25427d9d39d24d3cc6a9269b4303c +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..2f718b75957873eda59495eb16cc6da315accc65 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d524f173881e13c1d126f93ddd0c2ccb66a83b3d377b17a4fa6442b5e63dc802 +size 2589 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..d1227b44e579d0de25a07754285d92969784f40f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a48b7f2872e9c00a21834771d5f70680e36d0d8722c9243feefeb0f8559e85a +size 882 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..3d25d1673a3c1196c99f79af840fc33ff1a05bbc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fff579db089a9786d6daab0dbe8c25ec9046d9d5c59685732d5cec7bc2839810 +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..351dd6d8906e630fffc9fefe13c47bea1a153767 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ac281bcc3eb0e1577e1248f0d8419b0b731296f7fc203a7207f75020c468d4b +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/synthetic/tabsyn_n11_tabsyn_n11_20260520_234313/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/synthetic/tabsyn_n11_tabsyn_n11_20260520_234313/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/synthetic/tabsyn_n11_tabsyn_n11_20260520_234313/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/synthetic/tabsyn_n11_tabsyn_n11_20260520_234313/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/synthetic/tabsyn_n11_tabsyn_n11_20260520_234313/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/synthetic/tabsyn_n11_tabsyn_n11_20260520_234313/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn-n11-15215-20260520_234708.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn-n11-15215-20260520_234708.csv new file mode 100644 index 0000000000000000000000000000000000000000..b7fd5129b840c90aee81c62e71ab0b435ae92f88 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn-n11-15215-20260520_234708.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:079b489752f3ab1163d53b45a8d2019f1c91ff9ab50ad40e30f02a5e6b572f97 +size 1585519 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234313/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234313/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..15ddf94e63d5fbc7e81e00a1f0cd9cf5ad677c82 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234313/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20e55e61a63493b456b2051d3ccc2ba59efce75e22e2bcaa344fb40f467e682f +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234313/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234313/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..811e36fd50d99ed0a7ae3dd9bd2b9faa0c51ca6a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234313/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b215ec2f0352317d7423c1ce421fe154fd53a415e91536e414c8fe0988f07316 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..d2aa2a5117d35c4c935b25998e77fff6de8440b1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04e3ccc301808228438e15452980d943e2dc4fb907260c96dcb108e1310a0bf0 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..19a15cc7caef57d0608a3c4aba6776a64b07b117 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3c61ecd7b4cbef81837037be37e49d6a10fa511299d7378cf2c5d5be814c189 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..edbb66d5b252625129b8a7acf605c729c93086e9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b466da20c21fbc8f51821798db710f8d4f92dbdaa8332a7198676f63cfa781ad +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..b47da6e273e692bb9e8d7395005447e7d5f80277 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234313/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5639e3dc64c61abff262118cd4988b5f7da3b2803784dff309aab2ed1054520 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/train_20260520_234314.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/train_20260520_234314.log new file mode 100644 index 0000000000000000000000000000000000000000..f99f0a33e4c82a1e21c71e5ad0644b332fc4000d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234313/train_20260520_234314.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2efed04c846d540c1ed6986de19f06703d8d2235e2d7afe5a6b54770617b5ef +size 292512 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..f57e1c71b723a7d838571e1f7e54d5ee8bd0fb83 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_234826" +dataname = "tabsyn_n11_tabsyn_n11_20260520_234826" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_234826/tabsyn-n11-15215-20260520_234920.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..fdbecbb81ac19884d92ffb70ef898c40fe9a170b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_234826" +dataname = "tabsyn_n11_tabsyn_n11_20260520_234826" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 20 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..6d52b4bbfd4c70f6b3f66c7cc427b00b63e415c3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7626a856864f11bc07cfa27043b3ef5ff166c34bab7fcff9f171539a6b22aa39 +size 93 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/info.json new file mode 100644 index 0000000000000000000000000000000000000000..77004fd4811b7b306fc0365b7a69815ad565f22b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7284008e40dffc465c051deb379edfe72589b0223abbeb670976d4919012ee5 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/data/tabsyn_n11_tabsyn_n11_20260520_234826/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/gen_20260520_234920.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/gen_20260520_234920.log new file mode 100644 index 0000000000000000000000000000000000000000..df7ccad1e15d92856c893a25c87581bd82719fc2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/gen_20260520_234920.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c9412b1e071337ae5cbf82da9d490a0e39fd9248fc7cc1127bc2cec4e703a02 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..65fe761cccf4329d35d1f4a01bf28c7d8112d114 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84e18906e18d4695538a91adbaffd657dc45ff4e1fb1b039214044f7d9cbdcf9 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..38b2441aa8a5378c655037d824f3ff7a41373d44 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6398c01b3435dbba6c4a44d1650ce93a6b631c0feeeaf6e7d5529dd8fbab9da +size 2593 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..3cb293142e01ae8d93a234e97eee2aa9ab02fa6c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10b9ef0130c7cfdff4ef05cfd1a3dda7468a62ac2b6b9d61b1e362441f007f5c +size 881 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..77db698623f8dc7d17e7ca8454dad3fd02fe21cb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70bf3a806df2e9656aa06f9d13583cb32c9f4a17339cd80541938e8959b064ef +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c06ed3802b77b9ab78e132f458c94c7854732bf9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e09ff901e86a5c68b962829ab14a61ee8f8dde9b50b9a5d2119ae81b1e12076d +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/synthetic/tabsyn_n11_tabsyn_n11_20260520_234826/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/synthetic/tabsyn_n11_tabsyn_n11_20260520_234826/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/synthetic/tabsyn_n11_tabsyn_n11_20260520_234826/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/synthetic/tabsyn_n11_tabsyn_n11_20260520_234826/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/synthetic/tabsyn_n11_tabsyn_n11_20260520_234826/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/synthetic/tabsyn_n11_tabsyn_n11_20260520_234826/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn-n11-15215-20260520_234920.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn-n11-15215-20260520_234920.csv new file mode 100644 index 0000000000000000000000000000000000000000..61f5203c6125c44ef02739ef282d5f5f76f4e73a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn-n11-15215-20260520_234920.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98be322077830453f1a9c525cce4024a8080218fdbdfcefae8cd33ed21ba8edf +size 1585915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234826/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234826/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..da6228055e744921ea30ba0c4bc0d6c7a997ac2b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234826/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1161a2e157655c1fcf1e22c6b58ab92c6e03587f7d054318cffb3bdcdcf60078 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234826/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234826/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..56df489ba0beb2f427c14bb57a34bbbd2767b5fc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_234826/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbf4359d3b41fa1f7fb3a199d4cf18a045e0f7c4d85727fdd6470951da3f92c9 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..b1348d05fdac1ae66d1ed828ff8ff01c10f65fdd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd71ad2396a567a79916020f4b33af0fc854adfaf4be156ed510015a3bafa1cf +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..5137a7cefd5266b169c51e845340b9916bd79dcb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f9bb5c69b832a1fae3e0973c196e555ba7cb05f17cd067141fc2af047e29ce2 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..b57bfbe38e33c1874804cdaf5dbe6f1c48a5bcf5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e3107993226c6ab3bb3101d694e27f5f0e870d3e847f51d1264ddcbea0670ab +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..6400b030b3137097be2743fefa3cdf39b65e4615 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_234826/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ae956162ed171a514632bcc82baecb865389dac64d807352beea41f335cfc7b +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/train_20260520_234827.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/train_20260520_234827.log new file mode 100644 index 0000000000000000000000000000000000000000..13de3059b704a49cda1fb15c8a7f60bd41f00fa4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_234826/train_20260520_234827.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f463854897d53f275b59c3dcb25ae2351a8ff41a5425d57765a8cde52bb94ec +size 78769 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..7363283d50d943a5684f136e2824fa9535d4ce3f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_235038" +dataname = "tabsyn_n11_tabsyn_n11_20260520_235038" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_235038/tabsyn-n11-15215-20260521_000028.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..7b71b5b58e9a839263384b14f3df86471cf2d00d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260520_235038" +dataname = "tabsyn_n11_tabsyn_n11_20260520_235038" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 40 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..419696ac1e79930b45724dbf049de748da347918 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4af986b81a1c64b46c810c3ef668ea549767f5e5b321487541b75062baa17a8 +size 92 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/info.json new file mode 100644 index 0000000000000000000000000000000000000000..a9c286bb55be75b2380550a8272418c389b0facc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da2ed7939b3c0fa24d174575c677f443e207399b4cab04319c2e5fd8cb67190c +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/data/tabsyn_n11_tabsyn_n11_20260520_235038/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/gen_20260521_000028.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/gen_20260521_000028.log new file mode 100644 index 0000000000000000000000000000000000000000..a0b14d79d2e4ffb97de5ea91ebf1704eb2589c93 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/gen_20260521_000028.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3f55c8393d500576dd4670e1c288cbc219d09c2e2f250d8c585380a63de5089 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..92091b6c76039f49f66db7db134383378767c01e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e9f936ab2cb45d6c7c8a4a57a2dae9e1b151563a9409ff0144092101b5239e4 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..8b119cd3eea4adcd7e637355dfd1d471783290f9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6372c29a87e7a5909c00303d07ba4db15247999b4c77c0c7061294e37b419e7 +size 2589 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..b1c633e4242bda9d77fe3eee16169e5ffa9426d9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f66a8711198e8558f3cdd00c28838af53d0abd53831552bf8916eec1f5784de +size 882 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..8364984fdc093eeb29f49d55878d389005b0a3de --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df5eb59472918a9746c28280a95f504a42c88f826cb81d40959093b921ffdc7a +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..a3d01502a058ed14811584e595740be69856cd1b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fafa046e3f101169f2bc522637686ba4687cdcef05734be24e6348f997d3c680 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/synthetic/tabsyn_n11_tabsyn_n11_20260520_235038/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/synthetic/tabsyn_n11_tabsyn_n11_20260520_235038/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/synthetic/tabsyn_n11_tabsyn_n11_20260520_235038/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/synthetic/tabsyn_n11_tabsyn_n11_20260520_235038/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/synthetic/tabsyn_n11_tabsyn_n11_20260520_235038/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/synthetic/tabsyn_n11_tabsyn_n11_20260520_235038/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn-n11-15215-20260521_000028.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn-n11-15215-20260521_000028.csv new file mode 100644 index 0000000000000000000000000000000000000000..e561dcb7286f995deef10c5f1cac2a846977a226 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn-n11-15215-20260521_000028.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a523e8f36fddf2f78ff6afe98783fa63c8568ffe585f734fe8b9817e3263c6c +size 1587486 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_235038/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_235038/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..f938f472c6e11a85a7f69c6eb62174faf30eed53 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_235038/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e2abf82ed9144a7f50969f775e37cefe9a28c24793278840df7a5635f4d1e1a +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_235038/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_235038/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..b407cfcd60831eac0aad52f38f9706762b7a2ebb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260520_235038/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d0fdaa67988783eaf46a98a7fa404c7ca9cce554b64314946b20bfa9a39e0ba +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..940fa2643b506c63034aeda72f3651324a34a5e9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1effe02fdf5770d3b84288a7e52ed1dcfae953d78a2b5aede8348c3c4b0113fc +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..e7cec1f54f1ad5bd6007c533609271ded1af8559 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93e06ca3d485b679dfb85c4c2000e6afcd7c8a421dc3541366f6e279a622fded +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..cae3897283b98ad9f3c2f56b9bc6427c010e3e3a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36db380281bca5945b1e40c5fb4f8283602e0521d273d1d9462c9623f6ecc9fd +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..d6353bf0b6649dd7c35512b01e977b41b18b0de6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260520_235038/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54b69f23c0126c262aec5ffb863c380253b883b6ed2e539c67b77bfb570b1f7a +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/train_20260520_235039.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/train_20260520_235039.log new file mode 100644 index 0000000000000000000000000000000000000000..2e701ef2769c7dc861bafbdcfeec327d1ce240d9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260520_235038/train_20260520_235039.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a39dd3280b3144b4474b8602c61af2134b61a13287c58aa9fd5e73bc769fc462 +size 551193 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..c528034c2be7c01e3883f9391208abf8596e3e94 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_000146" +dataname = "tabsyn_n11_tabsyn_n11_20260521_000146" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_000146/tabsyn-n11-15215-20260521_000536.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..9ab14f10fda7778d85987786515adf5ba4fe84d2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_000146" +dataname = "tabsyn_n11_tabsyn_n11_20260521_000146" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 15 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..3514da489316d37fba8fa0473a2906083b55e225 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1809daf746cf1fdae2054219bae447b83a76f773b202c9e17293ac872d1f89ca +size 93 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/info.json new file mode 100644 index 0000000000000000000000000000000000000000..748fa85c4898cd5b2d15de84aeb377d773d37052 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5ac82f454910f85b150ddb05c997e8c311b811da446b2c86f7ba0244467d07e +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/data/tabsyn_n11_tabsyn_n11_20260521_000146/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/gen_20260521_000536.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/gen_20260521_000536.log new file mode 100644 index 0000000000000000000000000000000000000000..f04385e4e68ee9cfe30cfbd2f4a617f463d6c7b6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/gen_20260521_000536.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8156255f1dbbae7ff2ba6a67057b3d844fd6bb5ba8405fcdc7bb56a3178ff56 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e2afeaa68af7fb489ed8c71a5f4f8f7122fd4248 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efd2595b4a60426a3c394308879017f2b314acb492ff50dc20f343bd6467a0f7 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..c8bf56c27830799718870bb65a56228c0917e626 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c22b7c102c7eefa8e8d02109b7694c95f7d865663bc20c1b6890b0a8ad643b2 +size 2590 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..62c7e1d78b2d8357104335d775faff2ac942f2ab --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed271a47c2c2940f63ba59c99e1ef903133ad8a415186c381385153b8d2f985e +size 882 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..d611a59aa9a277500bd51949025d469162bc4160 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d99a4ee8b083bdf40e64eb7e6e76c4ef8b240080a1f706912a90c22d892aaab +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6af8699b9f3a773fbd99d356a8a9a54ba44e35e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de8ec091a3eb2cd3a7bac1b5df39fe928cbbb8e17d947af1f5dd7b39c11f44ce +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/synthetic/tabsyn_n11_tabsyn_n11_20260521_000146/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/synthetic/tabsyn_n11_tabsyn_n11_20260521_000146/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/synthetic/tabsyn_n11_tabsyn_n11_20260521_000146/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/synthetic/tabsyn_n11_tabsyn_n11_20260521_000146/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/synthetic/tabsyn_n11_tabsyn_n11_20260521_000146/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/synthetic/tabsyn_n11_tabsyn_n11_20260521_000146/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn-n11-15215-20260521_000536.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn-n11-15215-20260521_000536.csv new file mode 100644 index 0000000000000000000000000000000000000000..f71898971474cfb1ecf95374c008e5c8758a65d2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn-n11-15215-20260521_000536.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fe94cc54e6d4bb0866fc91e8b592e7bd40fd0719fccd14449434f60f915f217 +size 1585010 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000146/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000146/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..78e73eb13e3ea3ace53a23c630d2357a52708269 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000146/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6a1207f4b31216f1f4e044c0d65f90da4167e46956d24f65c77250e3b9d0be9 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000146/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000146/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..481f9b544583fb088a945cb69a3d92a5d1c09dd7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000146/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed33532c8fd295d98b9c4a71fad194c5fd15b19836e18c5eb1faa2713ae9472c +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..3120710dbe76682272c1a1f2c7318c2fe0747f69 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d364775edaa344bd070db7ee202601f3c53c847a0b7abe6b120d49803d3e5d1 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..71098d38c87bf1c2ebeb408170cc36cb07bc1e66 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd1b92d8dd3e209cafa9d8161495b7911df71d8358c889abef488fed8e12d430 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..ea2772511be4d2b753693c86b03e885c6eb82761 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9f0405e955686c10ea42b24a4dbec00f5763e34374af20f6cb539c8bb015407 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..03a6a51239ffe5cbff262aacf488f20305fcaeea --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000146/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57da0ef9bd878b8853566d8d972cb43e3d049b7bae9378099cb9222ed2d7c39e +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/train_20260521_000147.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/train_20260521_000147.log new file mode 100644 index 0000000000000000000000000000000000000000..a1c94b8a07197c7e8ad6fb0155c7df3587a1da67 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000146/train_20260521_000147.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e00783d551ecb934d8bb04c50e073d0b64c3dd9230ecd488029bce4a38dda689 +size 149386 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..fa108c60d1d7c72c25f0b0a45fc175c85e682927 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_000654" +dataname = "tabsyn_n11_tabsyn_n11_20260521_000654" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_000654/tabsyn-n11-15215-20260521_000719.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..dd628a669d945afdcb60d55e10a2651e20091ffb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_000654" +dataname = "tabsyn_n11_tabsyn_n11_20260521_000654" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 5 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..dae584c58e4d893e56e48b5a31cdb97ed858567a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b89d8d4964c3da2cfe8c4361205800fb222bfec69ba8a220eb230562dff7dfbf +size 92 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/info.json new file mode 100644 index 0000000000000000000000000000000000000000..0a786c1e00b96e607d691e12cdda87870ddafad3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b164f53c7476c7ddc917a1ba9404c1a22e3d64309a29015603ad10d77b1d88e +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/data/tabsyn_n11_tabsyn_n11_20260521_000654/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/gen_20260521_000719.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/gen_20260521_000719.log new file mode 100644 index 0000000000000000000000000000000000000000..93a39017baed9c3f626929dc894aedc4afa933bf --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/gen_20260521_000719.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:203b86e33bf2fba33b1980bed6804e857230a906bd604d8f30d71a3a898430e6 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9471579b693e714758bfe2418bbf382b8a01b7ca --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a4c45f0c4057909b7997fd291e95154fb61e0bd33f1427d9578123e9f2f9a02 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9261857afa7589e41341024deb8c48920de29a44 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93e523ff5f24a9750b41fa35afbedafed131c643990a46d8c665d951575754b4 +size 2590 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..278ab8e83cb029fd4917072ff3260cb84dedd7fd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:543c9bcb012dddea93adbdc33d2595399e7ca5b4e8bdc3ffafa2772619e7ef7a +size 880 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..618fd411287f2565830fa8b5f55e4255619192a7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9808962f2168952f6a263e18021012e2d4afc4a7cb45a3a8944931e0443dcf19 +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fe64510e669d57681e05eb452864dd69092afa45 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d85a38b3daf454e3ad57be36ab55adc53502302b629ae025caf090cd52c7cd8 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/synthetic/tabsyn_n11_tabsyn_n11_20260521_000654/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/synthetic/tabsyn_n11_tabsyn_n11_20260521_000654/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/synthetic/tabsyn_n11_tabsyn_n11_20260521_000654/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/synthetic/tabsyn_n11_tabsyn_n11_20260521_000654/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/synthetic/tabsyn_n11_tabsyn_n11_20260521_000654/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/synthetic/tabsyn_n11_tabsyn_n11_20260521_000654/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn-n11-15215-20260521_000719.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn-n11-15215-20260521_000719.csv new file mode 100644 index 0000000000000000000000000000000000000000..810cc956470a0368ce680f6b10b0a94103e6c39c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn-n11-15215-20260521_000719.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:479bf54b0b1ce1ddfe5bb69e5f648b4654f669d438596dc7bc029af01f785422 +size 1588022 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000654/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000654/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..ca01b72f4c0291174f0f8a632d15e9ee870d49c7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000654/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65d49ffa3917313cb2a105a0df7225df7c6d46196511fd8361768c9cf161c1fe +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000654/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000654/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..45fa38c9189b3ce68df138dd91e3efab1df3b23e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000654/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28190d7b6129db1bd43def9d0b7fdd0a49bfc5c04431a7c39089d664a19f0d4b +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..3ec32a52bf1e77e128c09740cc40625bfae091d1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcea8e81ffc39f272aed96cc7de1b55b42cdc425f2f33ab3284c8ce2f6119a6d +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..69be88138b1ba80b3431430c055965a7e13f64cf --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db53d1b9a7c96af93f549736667ab07572cb2f6fc597851a7fdd49526e2f03f2 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..bfd4fe092c8d6bce900886c9d6a1fc9653745628 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fea85dda74e126ca7af186dad21ed197e2a8f193e8e70f33739cde2d098bff7 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..2a5fe3220a0f4539447c6b034c0791b085a807d7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000654/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:534d43a4ef6398507ca942883884d8d0ed0823b97e06f390fa13168c634c3ad5 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/train_20260521_000654.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/train_20260521_000654.log new file mode 100644 index 0000000000000000000000000000000000000000..18cc06fb26b0e5c09a73fdcac927aa4db1271680 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000654/train_20260521_000654.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b3f1efe104468a6e0d6441611d137c61287e28687ede7d65a679fe250ff8f4e +size 68732 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..b0ad34d551b8b4eec25b1aa8d172f7fc5b622917 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_000837" +dataname = "tabsyn_n11_tabsyn_n11_20260521_000837" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_000837/tabsyn-n11-15215-20260521_000923.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..226e11a28a0f3d0d8d5c29bcabcd93ef54f37f70 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_000837" +dataname = "tabsyn_n11_tabsyn_n11_20260521_000837" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 8 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..7106d247230a91374f75cefe067461771f78885d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68beb2aef21200fc08e1f79fe9e100def30beeff30cf26cdfc7797805afc3d85 +size 92 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/info.json new file mode 100644 index 0000000000000000000000000000000000000000..314eda4eb316de9444c5d976fb099d2370d1519f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:291cac67a62305512d4f70ac45f1bd34a5ea5c67d717f2e18d8c9263511a7b79 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/data/tabsyn_n11_tabsyn_n11_20260521_000837/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/gen_20260521_000923.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/gen_20260521_000923.log new file mode 100644 index 0000000000000000000000000000000000000000..be0b2c323476556c1957bf62f0f3814e83f601b7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/gen_20260521_000923.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36424848ac6de0bceac21528cfc63b5bb80e82b981c18005f443469384599232 +size 949 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..cd76e3465fb835f257d8e77d506f8166229e756e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6c87c6fe9704fdc20c4a608fa1c3c57f146d421735cebdf216d35c73ff0bd62 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..e32b13700ac7aa1b48d402ddc4824afbaba72232 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8b6337e4d2efda135ba8d6279c1d574398f81856b1acc3937804c75957a5542 +size 2590 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..fa97a5fc7713d92d8bf9dfb2c8843b20c67770d9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26d9aef8bfd2d868fec7aec1ff998e7e87e38889de212d6d05facc0a787be4ee +size 881 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..2d67817c1ce86e0136002c405d8a6f2e93f9c7a6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e359c3866a8ffe8bc3f4f3c9ac757c71baa0ef8b27117e5f21b77e4e349dd2b +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0cf3d8a201903226439de574364a5ab6be679982 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc7cdaef7d970e5ba39641d9929fc35c137f2e2798c42909da348644455528ed +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/synthetic/tabsyn_n11_tabsyn_n11_20260521_000837/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/synthetic/tabsyn_n11_tabsyn_n11_20260521_000837/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/synthetic/tabsyn_n11_tabsyn_n11_20260521_000837/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/synthetic/tabsyn_n11_tabsyn_n11_20260521_000837/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/synthetic/tabsyn_n11_tabsyn_n11_20260521_000837/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/synthetic/tabsyn_n11_tabsyn_n11_20260521_000837/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn-n11-15215-20260521_000923.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn-n11-15215-20260521_000923.csv new file mode 100644 index 0000000000000000000000000000000000000000..db16b8929c2bde147d5b97409010e3390f241a12 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn-n11-15215-20260521_000923.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10460ff7352f56b734ebf79ad28f62baf5b467499ac352af005bb3e06285a805 +size 1586103 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000837/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000837/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..ae41c52f7f23720a208de87aa5d28c81d58e89cf --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000837/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:228247a68a2dce761e424667cfb83bc040176f8ed471da59556176823616b786 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000837/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000837/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..bfe34c9f07aadabf503863c45b5e0473f08c5844 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_000837/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:987ec01ccbc00a13a69b83839fe525324726048384253b32adcbfde3efea14d3 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..d225a3f6b8b0e7abaf85f629065c967210048c50 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fda2947f45c79020a173a94141a02bb1a376e0f990881c44ab5c4a33194b66f6 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..e61b94d292417feaa428ae7199482242a8278d45 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e970e5e816873cbf74bcbb8dc0d1f27b3beaa4a2ba49061221d52e42c334c679 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..a0d3cfe3688b0f6ae25b9ddb1889c517eb74ec16 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3dce14eeebed469ed183e45411c6d68c936730dbb20a99c493c502366de2c19 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..9b2c7e5e9f8646c49c2b9db6797d4179748ddef8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_000837/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea19b23b8aa143fc88772c73a3245e05dde8629f6ff8795a52ae136f2b8234a4 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/train_20260521_000838.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/train_20260521_000838.log new file mode 100644 index 0000000000000000000000000000000000000000..1512e56590252719d58b85b666080b4ece4a8160 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_000837/train_20260521_000838.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5da60bb7700c265248f20237a27ad4f509a1136607227c41b7a67140de154a21 +size 110392 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..9e98dec353a1cd389ab182915444e179a6210b39 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001042" +dataname = "tabsyn_n11_tabsyn_n11_20260521_001042" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001042/tabsyn-n11-15215-20260521_001211.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..3aab061e8df5900976835685768b89b9e6b8d35e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001042" +dataname = "tabsyn_n11_tabsyn_n11_20260521_001042" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 40 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..458ddf4177c9a033922f2b19537b9fe55dd1c6e4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:347343eaf3383580d98996065b97b7165d3d2d041c08ac76c1433062cfd3588b +size 94 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/info.json new file mode 100644 index 0000000000000000000000000000000000000000..d78c97dfbec608d54161c733d11d1f828f5091d9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:505cf8655fe9b14a4a618d2a23ffbb7161e9e3dcbcf302e9dee9a4b6e152a1e1 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/data/tabsyn_n11_tabsyn_n11_20260521_001042/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/gen_20260521_001211.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/gen_20260521_001211.log new file mode 100644 index 0000000000000000000000000000000000000000..2da8c79d0b4faecb9d526ea64bec09fc1cedde8b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/gen_20260521_001211.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0fd3365047169a9f564b8002f2122381b8a5cf2021fb9cfe07e36b73255bf4d +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..019741ad37c52d93d9050a676411fc9aa36c60da --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49bae4437738772e4d8920f12ea4b890cfdf0d318672eef0b3bf225a93c9312e +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..0bf7e8f936c8f5d2e33c3deac2a79e76096f3d37 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7eb7e5edae1eb21dd585bc823618dfc3837c699e06b5e07b20f0d59e6864054 +size 2594 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..bcf1df594722f979ce2389a5f3b637231a21ab76 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e846244fd9f329f5a45a673bcfe3695ec06bb2bfb42be5e35d974b046eea0264 +size 881 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..1bee4fed357d058c050d5367c0d3bd655340efb8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69b12ff67f96f3bf36c2f9c06baee81e62642282077fc5c87905a11175103c2b +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d499690ed79c06e1a9fb360f4647e362ca65e070 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:274b7ae19f7d46a6b846e819afa0ecd6b0db55ccb540af0ab507e9cd6c87084d +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/synthetic/tabsyn_n11_tabsyn_n11_20260521_001042/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/synthetic/tabsyn_n11_tabsyn_n11_20260521_001042/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/synthetic/tabsyn_n11_tabsyn_n11_20260521_001042/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/synthetic/tabsyn_n11_tabsyn_n11_20260521_001042/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/synthetic/tabsyn_n11_tabsyn_n11_20260521_001042/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/synthetic/tabsyn_n11_tabsyn_n11_20260521_001042/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn-n11-15215-20260521_001211.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn-n11-15215-20260521_001211.csv new file mode 100644 index 0000000000000000000000000000000000000000..bf6f409a753e7709d3cbbba8ab6a823410777201 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn-n11-15215-20260521_001211.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2584ad3fd3e619ee3fab5aea4ae58a509850ca825d3b2f44030684fb857f3a6e +size 1588096 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001042/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001042/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..8f2523fddb5e641a2683b4e96c9db261847567f2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001042/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1d6c84778e45c25b6be6dd2f8914eca100b327689ad9d356c152efcfa040765 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001042/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001042/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..38c6698f549dfbb81d2da18a9c72ff9bd55dc9c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001042/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3a8c7846e3d208b0448a9f4a75a6036c1ef75d51a1bce1fa0b169b89fae0795 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..dddeb0c3f0696540de68727bd9b3036b46308dd6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:126efb34bd60201d139c0ebb625995bf5ce245c57fd38b2c2be8f5a0f69f32f5 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..f082a1b0f8b61768fd8a4a9b118144875fc8ca53 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d7d6c81f2c746780b57bf449c5df249f4a95a42c989df9d0486cd02365e4183 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..8970664124032c8c72f98149008282a404a77be6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d75ae05617c547bf01c5716d398fa45cccff4e7c05aa53469b00f3ca3a0cbdd +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..6c016d6f094347c5efa446f5f2714a0206891020 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001042/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01f94991ae44cd91c3a4a30482a14b0bed00602ea91ff7d320bf578d3caabc54 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/train_20260521_001042.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/train_20260521_001042.log new file mode 100644 index 0000000000000000000000000000000000000000..5a01dcdc95bcfc69bcd9bc8b094228c4cd66ea83 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001042/train_20260521_001042.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83cdda4aaa82544153066deae853237740de18655b20984aaaa4cd7a8884c3e1 +size 103303 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..8f6b1bb7f348a1f9e716b72840f232312bf9736c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001328" +dataname = "tabsyn_n11_tabsyn_n11_20260521_001328" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001328/tabsyn-n11-15215-20260521_001613.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ef51133009debe80b72189a7bda39c794b0dde6f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001328" +dataname = "tabsyn_n11_tabsyn_n11_20260521_001328" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 40 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..c83395389175134ea752c185caf434576daa4652 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c33bb0e37b1a5fc4ee988a4cf8754e833b3d6794472b5afaca855f89ca4cfabf +size 93 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/info.json new file mode 100644 index 0000000000000000000000000000000000000000..a47922c2d9833754f3070fea22cf1705f1581f4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:782b2991821afbee1ad6513557e343ee565cf190fb573c79bea14444add06d1b +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/data/tabsyn_n11_tabsyn_n11_20260521_001328/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/gen_20260521_001613.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/gen_20260521_001613.log new file mode 100644 index 0000000000000000000000000000000000000000..a0b9318fd27971cd731e3f79dcdb9e801e26a70a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/gen_20260521_001613.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b5c7afce9342883a72cd4a327dcc1f5fd1559cafdc8bf3c87691877e1a43908 +size 949 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8ed2a228b3f652ca0ecb9b331234f536755ba7c0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aae1d2d56935ed02d8f99a6901e7e20f82645aa1019505dc64623bd785222078 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..7334352a024d90bf9f006c6acf39afbd04a3669c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9bd86e612ec466e62e43839ecd894a6edc7db47a682b6f89a6872698d9d09d4 +size 2593 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..093c1ee34e72bb0bfccb8f6ff4206c1b4d2192e8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bdf40f8878ec2810843ae476e95778226941ff8b14cc3d24cd49581bb60aa93 +size 882 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..af1a45da5c179d7498e54a1d0e185ccb29676d2b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e874173ccb198ddae7eeeb07776f372ce7d3e9e5401e6537a29533446a900487 +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e82fb68b7c0e5ee050e340ed88764bc8006f7735 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ba1302575d1559cb8c667542e035d3591e9cf2ebc3adcd273bffd1724c0acc1 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/synthetic/tabsyn_n11_tabsyn_n11_20260521_001328/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/synthetic/tabsyn_n11_tabsyn_n11_20260521_001328/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/synthetic/tabsyn_n11_tabsyn_n11_20260521_001328/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/synthetic/tabsyn_n11_tabsyn_n11_20260521_001328/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/synthetic/tabsyn_n11_tabsyn_n11_20260521_001328/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/synthetic/tabsyn_n11_tabsyn_n11_20260521_001328/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn-n11-15215-20260521_001613.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn-n11-15215-20260521_001613.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc2ffe1775ef8a10ff115f4aba2b619a2713d9b6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn-n11-15215-20260521_001613.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:209752251fb97586d50bec442faeae65730012bba85bf6056e8bec946b39ac09 +size 1586316 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001328/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001328/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..9ee1c37dec77e761c820054d270bcecf55afbc44 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001328/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ef9f05e561887701b3490686813d3bf8a96a940cda534805decf6bbb0db0505 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001328/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001328/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..4b69ba86e5f6116d3cb9bc26fada7be447c67b79 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001328/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9428eaecf64551c4f785abbc23b57540f2fec3fd778acc22a2834b158284402 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..74e5a52bd3ffca51cbd2d451e6b3ae24dbea4590 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2386f5148181c203c075819f44e8d2b5311d595fb7fa61ebaac3c5f32b1cd875 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..ccce0fff3093f3a2a4306d531416bc7c1091343d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a81e841bfe831f3f85a0ce2e321667730239967b9fc22863a0ec93578979fad +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..4c25e847090dcc40cce2e4b58454381fec160b1f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ac65ad113c8e2a1373df8e312db8b42a64ef9dc55941f4113232cecbb541c4b +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..302e001bcc00c34a2493cf0e6655032e3e2e3f15 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001328/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ab5c10c369e783960c141ba51928cb9a5fb13c8fcad6d048f5e34afdfa11bf0 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/train_20260521_001329.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/train_20260521_001329.log new file mode 100644 index 0000000000000000000000000000000000000000..576efed1255fb5046b5d1d040e7ae28304140e45 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001328/train_20260521_001329.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7444117719d87c1eee5e4b5cfa74c6224e0319612e6d51e0ecfc4c8ad9e2a7cc +size 298166 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..47213dfe7d3f19dce2fb546d4a281bf78567612f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001730" +dataname = "tabsyn_n11_tabsyn_n11_20260521_001730" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001730/tabsyn-n11-15215-20260521_001818.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..affa0ee3f6d917408214578b06dc248cbeac7aa6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001730" +dataname = "tabsyn_n11_tabsyn_n11_20260521_001730" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 15 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..1ddda0ebb2708be02d0f212e357785e4fd362eee --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6c8640228530285becfdeff638e805ddafb2f5e5778d927f4df4103695fccad +size 93 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/info.json new file mode 100644 index 0000000000000000000000000000000000000000..1f03424916deeaba22325df37db3c258223419bc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f94c30d5609c782350a1ed7504ffa9e183e5a9459f371c565c21ac9f3e2e04cb +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/data/tabsyn_n11_tabsyn_n11_20260521_001730/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/gen_20260521_001818.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/gen_20260521_001818.log new file mode 100644 index 0000000000000000000000000000000000000000..1cc336c6a803525177d5687fa82b26cd854bfe02 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/gen_20260521_001818.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fc523841cbf48e8e270f02481d227b5850924c1e9e5f4140122d991da07a169 +size 949 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2486b5c5099eea73d060d839cbcb628d72b3c02a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:febac6d6c3a3bd97d50f6ac634c8924fa9a46b19e58326bf7e6911d9137be915 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9d980a0ae3e2c5dd808afc12d1ac4268bf011ea1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79852678d5e0502cb67d23022ed4d0992c65eaa800481a359abd8a8ff8b15e92 +size 2593 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..63a2bddeff6b16749c72f55c1a592ade95381891 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30d106fc4c7e058dcee892feac27e2c48d6b271487edc6a3145d36dfa081b07b +size 881 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..ce15c86d2ce5c572123469f6f5bcbd5ce8409b33 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a24dbe77797caf5f6edfb0d789bb46cdaff7846ef74045bad95a370a8a38f8c +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..13e95d22a78768299e760d416a9523c85ae17f16 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32d421290c4da72335d6d9ee86d02262aa3d2888964430c04d9887ac086d1857 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/synthetic/tabsyn_n11_tabsyn_n11_20260521_001730/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/synthetic/tabsyn_n11_tabsyn_n11_20260521_001730/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/synthetic/tabsyn_n11_tabsyn_n11_20260521_001730/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/synthetic/tabsyn_n11_tabsyn_n11_20260521_001730/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/synthetic/tabsyn_n11_tabsyn_n11_20260521_001730/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/synthetic/tabsyn_n11_tabsyn_n11_20260521_001730/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn-n11-15215-20260521_001818.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn-n11-15215-20260521_001818.csv new file mode 100644 index 0000000000000000000000000000000000000000..e85ea2e1fbbb38a62b0d2bf40a95b8ec1fe8aaaf --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn-n11-15215-20260521_001818.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f59d160b9a07bf99ab4a5fe89ef11d6011c386bb9347cc99ea07bac2f21cc304 +size 1587518 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001730/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001730/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..cae3b7068dc649ab750cc605d7e00af54b4eea5e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001730/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32e5ddc85a60bc35b5fb75fe3802ab6be3afbfc6a3ca17aaf9842e33d21bfc82 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001730/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001730/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..ad525543d97052a808a08afaf311b3a782e3db46 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001730/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3436a5c0e5b345c5a0588172634b10782ae266fdf0a474a635557f55868e0d9c +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..7a01a12c9079c00f5c3d3262299c00465923c68c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51bda893db0664f1276b371d80817efff26590fce50da499e352ae506ac4996f +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..dd0529034901c68bfcc0d66aeaf4656ad752a520 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:451e78fec99fb1631a4a74b30a3679ad62e8b2c47fa8dc603312fcc0a19c79bf +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..02670a97e0717a768df5f5731ed97757bee2c862 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e022e10a13a1f093ebad1026c06958fa68487dfbca4e3d29a582b897d6e25dc4 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..0025616e6b6efc03ba6c5f6268db9315ddcb050c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001730/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a0706befc204e304e2eda580aea91090bf79bd548ddd9bc1fe79527e81a3cf7 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/train_20260521_001730.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/train_20260521_001730.log new file mode 100644 index 0000000000000000000000000000000000000000..4c629f2974daf8c35de208ed475e0761f5a12aec --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001730/train_20260521_001730.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:637fdd48b9962bd7cfb86b412fc03196041ed35f6ab78c88fed5754ea07e1847 +size 184955 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..d622d73e4ea8ff2a4d9caa649e348d9aaf83d934 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001937" +dataname = "tabsyn_n11_tabsyn_n11_20260521_001937" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001937/tabsyn-n11-15215-20260521_002102.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..618ce35aa050ec416d0ffb2bd62e321b957e6486 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_001937" +dataname = "tabsyn_n11_tabsyn_n11_20260521_001937" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 5 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..001d57a6d122ae690ec69adf03fc6d74d2f7f20a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d9274f40a4f125eb8b282d4ff6690b7aaeb6931c07307ca3eafcd5ebea434a4 +size 91 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/info.json new file mode 100644 index 0000000000000000000000000000000000000000..1b8b9112c55fc01a0a6e1203c8cd7a81bc60e85c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35961dd12fdb45f8d3bff48a0eca117b2186e4269da60f897d14d4b59fb8c5c3 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/data/tabsyn_n11_tabsyn_n11_20260521_001937/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/gen_20260521_002102.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/gen_20260521_002102.log new file mode 100644 index 0000000000000000000000000000000000000000..afd9d4a07d6dbf28cac81f9ceb90d9ef4a43745f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/gen_20260521_002102.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7c4641e0f9ff2142540740e6308a76e1daf57be42659b290ba30d1f2f281d81 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b5ea68d8b442d199cc9d2a5177434f1437eb46e9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab164cc6c0d4fc5ffa8036f68e645935f262f06f11428d93cba0710337e7a7e7 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..ae34eb5b85eab5d4b1c316d61d16c4b843592520 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d6665777ab8d9fe65af9a53ac60df545c2653ca3e935ca8c8bf8df3ddb38183 +size 2586 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..28cadf42047ef6a6ed92ee9117f031c783f77613 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f7c491ca9979be34dbf8ec0562dd8c8905e906ab5c854197c4f135aae9b6cb1 +size 879 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..417c08ac86e01c3517223df19eaf1f61f32928e2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a61d1dca6b53865270b288c29ac5d9d1a0daa62113c6d21275194ca72724bfef +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2408f9bb2b50828b6016143c205147f276413c8b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb526a4802a3460f413fe0f65398463bcbd8c06537908e66dc9b589268873084 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/synthetic/tabsyn_n11_tabsyn_n11_20260521_001937/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/synthetic/tabsyn_n11_tabsyn_n11_20260521_001937/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/synthetic/tabsyn_n11_tabsyn_n11_20260521_001937/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/synthetic/tabsyn_n11_tabsyn_n11_20260521_001937/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/synthetic/tabsyn_n11_tabsyn_n11_20260521_001937/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/synthetic/tabsyn_n11_tabsyn_n11_20260521_001937/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn-n11-15215-20260521_002102.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn-n11-15215-20260521_002102.csv new file mode 100644 index 0000000000000000000000000000000000000000..a8dfcb5b342547a89a12c7c5d4649949b28840dc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn-n11-15215-20260521_002102.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48079e90eecc1d86e5143a767760967244fa8fdf5f29151536d2071746c6ba3e +size 1584354 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001937/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001937/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..4939034e0285da2f86bdd341d82574fa0841f19b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001937/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c96c33133fd3265b418404843293fcdc089cb9b3583cbcc54abf5a953a7b2ed +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001937/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001937/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..07c2b5f7a14be686dfce244f0ef365896a901236 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_001937/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94905374a5c44f27a7052e555a3f69b36deed330ff6be9244c07dd3680604436 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..3096e113db7d08d99a25172d6733293b09e3bbdb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32b95a8cd4136bddc03daa851ecf2f6fe7f826b8fdf1e06914a8eb5a7ba11de1 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..7d0318dd818f08fe7b14074d260738c22f2a7f62 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f7e880e77f6658f9b75c2d649648d8ef73bbfae3696720454aee01f6e94e5e4 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..d974227d525e30d005cf35e4f04238d74fa9ba87 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0acc86a743960aa43e17998c4dccd7bc1aa60dbaf90da4c31223b0aec52e29a9 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..f968663629eb41df143caff453d7a3ac3fdad594 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_001937/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ed0be753fb70e84f41da3692998e25e6de939b6799e472a415ac83cb1e5d8f5 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/train_20260521_001937.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/train_20260521_001937.log new file mode 100644 index 0000000000000000000000000000000000000000..b35b7f7716e9da26d942bad26d18d72a640d9418 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_001937/train_20260521_001937.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0dc1af0f83e7aea9692223095a5d073c7c9dba26ab6de115a5d5f6391b5d93ad +size 73415 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..35d55e7ae9ec18f068924c01878af321bce770b8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_002222" +dataname = "tabsyn_n11_tabsyn_n11_20260521_002222" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_002222/tabsyn-n11-15215-20260521_002324.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ff5689ef3e8f3c548fa07039cac24b85033ff3d2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_002222" +dataname = "tabsyn_n11_tabsyn_n11_20260521_002222" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 12 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..a085083049ff3d819342219216ce02a113f6d555 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f61f719822128043e56dbf25bc89f361fff04893e56203c8f2c0acb1de2d833 +size 93 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/info.json new file mode 100644 index 0000000000000000000000000000000000000000..1c13eec9f4291fa47e5ed595bd70bf5f97b5f154 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80c2af6f1b04fc3b40d869eb124843b2fefbfd3d78bb47ea9733f38f1c3a9ff9 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/data/tabsyn_n11_tabsyn_n11_20260521_002222/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/gen_20260521_002324.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/gen_20260521_002324.log new file mode 100644 index 0000000000000000000000000000000000000000..d0af284f3892133ffc169b699ccb6200f392fee6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/gen_20260521_002324.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cce4c04f6ca1a44525807dada5517285ddc8d15fce40419e8a62c220717df8e6 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d7dd6d7ae5974b0270ac2d36b02aca50a2eacec8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:334d263ddf0b838a8c629651df4fd042586fb8bd3846424da5b48837a208683d +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..3f39c6fce77bf903ae9b994ab3c611697cf3885d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:310485846602e77be873a1d2d3387cca60d22fb743704b60fa7f7f1be72e5e64 +size 2593 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..01d54cc0dafd188c0f4cfda521a7e01d9966e5b4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d5af5e15f884babe6a986887ad1b9cb41869a18db881cc622daed2ccae63bd8 +size 881 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..85c0e6c293b182fb33c2f2532ff71c9071111516 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2a8036f85a6781b9ae4d690404df3d7967ca5bf7f24c231ab3cb4968c70be5d +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..21ed9d259980a92b5808828086003dcbbc24a848 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b73448b042579fab121d29a1ed25e652ad05d1fe6d06bc4764767274a57bb79a +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/synthetic/tabsyn_n11_tabsyn_n11_20260521_002222/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/synthetic/tabsyn_n11_tabsyn_n11_20260521_002222/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/synthetic/tabsyn_n11_tabsyn_n11_20260521_002222/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/synthetic/tabsyn_n11_tabsyn_n11_20260521_002222/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/synthetic/tabsyn_n11_tabsyn_n11_20260521_002222/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/synthetic/tabsyn_n11_tabsyn_n11_20260521_002222/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn-n11-15215-20260521_002324.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn-n11-15215-20260521_002324.csv new file mode 100644 index 0000000000000000000000000000000000000000..f6ebf001f8522dc7756d5bb9d9b029f3b2e9c227 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn-n11-15215-20260521_002324.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98a73e5fda1282e6fdfe12d8ba2e726fa46abcbd7e85c140a10e79ebc6ad24ce +size 1587925 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_002222/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_002222/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..338e9f45d530f8a2c5f44040776bf41e7bc66c97 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_002222/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a175b86183037eb2b1a840fad2aed2907d51314aad2f713a30e7cc5fbf0eb3e +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_002222/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_002222/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..dab914a8353a6047c690b397d75451a01d374781 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_002222/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a69dc544ba214e369b08ffc32c6f54ae2c1349285044d21ec6381f3145edf89 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..123763eeeec36e9874aa2b6ae4bd90c78cda0757 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ad5b9301dc86f9d5e5cac0a953266fdb16eeac5bd349bf3f37403a8ff45826a +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..c5243d072683fc379929f98fe4d3610e3e0edf68 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc4e22cc036f347f6cb4eb07b31e523c671fad15ae965ff49d6a74dd792b61b3 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..6eb596eea20f87987dcbc9048b27b645f6fc47da --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6a0d8f09291e667fe209df336c8c1e047e804c5438cec0cb2de908fd40c7edd +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..bcbb690b8a0a3cc890a8b2d6cf758031c0b9306a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_002222/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac8a4ad61f7641fc3ee4f4883c125383e0fb5fc343f7903fb771a5db76e103dc +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/train_20260521_002223.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/train_20260521_002223.log new file mode 100644 index 0000000000000000000000000000000000000000..b5f561858ab57e06e4942229b1ea9fe3c9bc0999 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_002222/train_20260521_002223.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd8c2742aedf0718b214ceb86cdb9a6486169047303508767c03539d74e40209 +size 96206 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..2ab252f2ee4dbe5ffab11549052962c847459e48 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_034608" +dataname = "tabsyn_n11_tabsyn_n11_20260521_034608" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_034608/tabsyn-n11-15215-20260521_035821.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..3ca64a03fe7ef44fb5e6eef37b1b3b7315b8d4c9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_034608" +dataname = "tabsyn_n11_tabsyn_n11_20260521_034608" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 50 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..90785436ab9875fdd7d6645fa22ed221b5dab292 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45401fda97460f7f685fed22508a675769c9ee1670375dd305eacf37725733fc +size 93 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/info.json new file mode 100644 index 0000000000000000000000000000000000000000..c499fc57865d04106b85d0d4273229e17a2fd480 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2825010fde354c679597999572da188b52938f94eedaef361a47b8447fefb2f +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/data/tabsyn_n11_tabsyn_n11_20260521_034608/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/gen_20260521_035821.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/gen_20260521_035821.log new file mode 100644 index 0000000000000000000000000000000000000000..db97b550539cc2208d7db2622d89ed00391045e7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/gen_20260521_035821.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:818078c2256f223cadfdde4367911a506e8d46a85b959b8c88ed55fadcb94100 +size 949 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c913724253a6ec44cda386777c869baadacd58fc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:221ed8308f179b1936b7d2a887a3c2b4f6b8ee477c9e71fd5d5b66ef4b49afd7 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..17c599f68dce5432135ff4026bc5fd022f8a01ce --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e0eb215e8b5b7b3b463bea58b3185e04d010eb292ccdb74b121470cd3a19731 +size 2594 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..21b29e551de3a70c25ef305571df4f92a1b9bcbe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5e86b8bfe5355122c5a912a0c66589fd177d050758c6e5edbef28b1ae5954c5 +size 880 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..7add3f8ee95ff908393e58a28cee2c03bd611ef9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:579841f047a2212a90b152ac668c9a0c6fefa277f153b5cf05aee7bf7c8783fd +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b350172a06b831cd7f8ec66ece5a540700ae9026 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4276e2637ed7c4e8e5c47f891805bd9099f1c2f220c38427d9fe48fe8bb5fa3c +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/synthetic/tabsyn_n11_tabsyn_n11_20260521_034608/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/synthetic/tabsyn_n11_tabsyn_n11_20260521_034608/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/synthetic/tabsyn_n11_tabsyn_n11_20260521_034608/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/synthetic/tabsyn_n11_tabsyn_n11_20260521_034608/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/synthetic/tabsyn_n11_tabsyn_n11_20260521_034608/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/synthetic/tabsyn_n11_tabsyn_n11_20260521_034608/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn-n11-15215-20260521_035821.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn-n11-15215-20260521_035821.csv new file mode 100644 index 0000000000000000000000000000000000000000..5af58c3e51498c7aec6d12fa83766358e88a26fc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn-n11-15215-20260521_035821.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b7a84eef1df0fe5c8cf69c960ced126d7ed6cfe70e4c97d03f52ae21676e492 +size 1585861 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_034608/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_034608/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..09ad4906eb0d8285e0861c5808ce6702d18fa843 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_034608/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9041bacea1591336ecb3487de6aa14460fb726a2b65dcd9ddd7c74e736c05f31 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_034608/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_034608/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..fa0c6ad601fded1146d18ebc2f06647b39747aff --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_034608/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c8a3e2973ffb5ad0d8da33013c2a6298fa876e0e472ff369b41c9a1a59808bc +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..636ae8040f4739080d40fc7189365a0f4572b209 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc7386c4d882cbf1b1afba3c5e278a8e2d62a0343a22aeea83d924bba991bf66 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..0f5e3048e0d420f388b26c4d0108569f8fa99c86 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26566fbd5a5626938e80aeb0d26c7ee7cd4ffb6f106cb0eff675bc6b682e848f +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..d84c5c137ef3a81f5acf645959cff4552bbf74ed --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f411255cec55ba6d46bafccc4692a0173f387bf4647d1943cb004568e1314687 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..323c69afdca33f583a05a3e0e4b1851a2a70d07d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_034608/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cdcb707284842a46381f8d8b30e9f35dd4ad74177028169139673de0fa74a50 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/train_20260521_034609.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/train_20260521_034609.log new file mode 100644 index 0000000000000000000000000000000000000000..2cb250cdc0fb3c83c635e315ed66000c0e83a9ca --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_034608/train_20260521_034609.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9687fd0227dba8defbd28c48d7928d02a26bf3db4d5ff620022ffab77e23050 +size 499562 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..e5958e283e66516fb4cda142aead8e39f1a9ffea --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_035945" +dataname = "tabsyn_n11_tabsyn_n11_20260521_035945" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_035945/tabsyn-n11-15215-20260521_040942.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..fd1a111da05013d8b4d292f2a7d8df3df3b6dea4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_035945" +dataname = "tabsyn_n11_tabsyn_n11_20260521_035945" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 40 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..80f05b3385743727ae9469aef25e0b37fed9584c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60c8862622a1ddb85d2f5035e1c4bb6c56536f90de679e21cfefb5bd1315b1c5 +size 92 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/info.json new file mode 100644 index 0000000000000000000000000000000000000000..690d230a79d40a46cf01125b0d339da59a79288a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:295c78c41520590d4c1518f1cc92f5b66a25f27cc49dbfe0be3a80eee658893a +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/data/tabsyn_n11_tabsyn_n11_20260521_035945/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/gen_20260521_040942.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/gen_20260521_040942.log new file mode 100644 index 0000000000000000000000000000000000000000..c1c1a753185baa728a3e0ada94a327f7727daa5c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/gen_20260521_040942.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00240d44d430117f86605de93cf0257d30d202ef17d323e84b601683600c4e98 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8194f4698af855214c01361acef1ebdc3accf2d7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30efd52e9afedbed8eb33390be96693a3ca2117b518a6ace859c08ada2787301 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..0be11b1e40243964775228b5861690178c7ca27c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30f2a865bfdbc871917c6ea20f9d29130de9a78d6657dd3afeeb00fce42d9f1d +size 2593 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..577de5d17fd870348c72c9d856e70d4f3d5ebbac --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:255bb80bd3ad8eb21f2f79dc66a7fad5524500bb37773520f1119cc7aa197ef7 +size 882 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..81aa6182a6eb9bc61fe2076de6e016fa6253e354 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbff0c2a72236979fbc496448b0b26f18ccf4e105c6057481013bae8a487b92f +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..37fe58dda431a4a4c5439028c8b59a053c242c52 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a8a40c6daea6c2b82899618bfc2a38de2d531431106126d06e386af181a3e9e +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/synthetic/tabsyn_n11_tabsyn_n11_20260521_035945/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/synthetic/tabsyn_n11_tabsyn_n11_20260521_035945/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/synthetic/tabsyn_n11_tabsyn_n11_20260521_035945/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/synthetic/tabsyn_n11_tabsyn_n11_20260521_035945/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/synthetic/tabsyn_n11_tabsyn_n11_20260521_035945/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/synthetic/tabsyn_n11_tabsyn_n11_20260521_035945/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn-n11-15215-20260521_040942.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn-n11-15215-20260521_040942.csv new file mode 100644 index 0000000000000000000000000000000000000000..bc66e8c2ec408a87200f091dc0726c9fb97fe354 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn-n11-15215-20260521_040942.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:715a55caa40c4a26d11faae33fb14aaeabd59de84b894c75dce1cbf716188ee8 +size 1587623 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_035945/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_035945/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..a858a375f9c44ff1092d3eb865f7c79e95294e9a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_035945/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f057db0338d5719205f50301371758b4e89cb5a56fc44a9961c0d850df7e0e4 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_035945/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_035945/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..46ae8100308425091818806cf2fc576af5ae572c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_035945/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c7150745e1fcaf6106297e186518e5c9742bd7ee7961a7896353e46205b55c3 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..6e695f0763680b6f14570b6953e8440de0b80adb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b20504db830a3e08213ed10aef43c99ceff3fb811d794beb605e964e32d081f +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..e1493d6d611d51be8324478d4853f863d9dfa2ae --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35e669f7bb22b6c47667f4c905ea95703c4fa173741fe3e3d21c9af4e859b739 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..f212f874aed4e373eee6e4ef5f41d5efb69a3701 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff1d3b14cf9cfbf79924c1551f71d19dcd49e625f2e62a71f7a4b15ef0cd60e0 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..74d8841b15756edd89c9bbd03ca18339186c9076 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_035945/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b46c818448bab1899780b9d1a27158943594cd6b6d144fe7c7b6396ebee9cd47 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/train_20260521_035946.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/train_20260521_035946.log new file mode 100644 index 0000000000000000000000000000000000000000..5c1c1a821686606003cfdc730050d6348dca93b4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_035945/train_20260521_035946.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6aa8cc47cdceafd40d818bad6aa853ea18f1e1f02748a6e62475e67d96d81489 +size 448962 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..775e97e7295f7a5683ff50a9967a2747a8215be1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041059" +dataname = "tabsyn_n11_tabsyn_n11_20260521_041059" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041059/tabsyn-n11-15215-20260521_041229.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..647d1b3593d79c07accb6c06e2fe8b94418a170b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041059" +dataname = "tabsyn_n11_tabsyn_n11_20260521_041059" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 5 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..e0c1df50590f5d803b5fc3ca1d0b59c003f6b38f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77c05d2fa63694f2f712c85430eb6c828328901aaa5a688db28508dd54ea886e +size 91 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/info.json new file mode 100644 index 0000000000000000000000000000000000000000..d6161a83147918afaaeb3b246a233f94aa93dde6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa83e9d51b94e9ee33ae73e27252f9057aa4ff6adb3d65f7fbf03dd37eddcb31 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/data/tabsyn_n11_tabsyn_n11_20260521_041059/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/gen_20260521_041229.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/gen_20260521_041229.log new file mode 100644 index 0000000000000000000000000000000000000000..c85a44cdafe9a2a6b3279f0f00023547ff0b687b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/gen_20260521_041229.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03ca0fd157e980e755f18bc2c4951e5b57254eaf15edf080b55289fdbd89d563 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9fad00d4ddcd402e06aa00e909d7091127da5b40 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8aafc33e67861a34659b0f616ca547e0434fd00fda1db836750454062380bd70 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..33be1f2c726e2415297ad541be9401c66672ed7a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1184824af035fdb3de5c4520db31c57b44b958f140f4a51c5fbd6b1a7f10a589 +size 2590 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..5f8f861534009c3c9a2a6c20a5b7b308d9cd60a5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09896c38806460cc7c9c7f5bcd3a1708864256cedf94acfaffcf5a194c4026b7 +size 881 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..2b717c4c878a177e25526d1db77771af19636f90 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b145d9f55aba260995f58b33d34a284a73071f927de9f974044069ac6e2e4f23 +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ccf44e8fa6495432292bcf11de4acb618b6e51fa --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc6e3f844d3d0398873b2fe6d8799db7b97ffa0cde54a5e43743ecb0149cca97 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/synthetic/tabsyn_n11_tabsyn_n11_20260521_041059/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/synthetic/tabsyn_n11_tabsyn_n11_20260521_041059/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/synthetic/tabsyn_n11_tabsyn_n11_20260521_041059/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/synthetic/tabsyn_n11_tabsyn_n11_20260521_041059/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/synthetic/tabsyn_n11_tabsyn_n11_20260521_041059/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/synthetic/tabsyn_n11_tabsyn_n11_20260521_041059/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn-n11-15215-20260521_041229.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn-n11-15215-20260521_041229.csv new file mode 100644 index 0000000000000000000000000000000000000000..468ddf73e5b476280ff8c60dcd57541fbb3091b7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn-n11-15215-20260521_041229.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:535869a94ce3b8728c78d08ad968078b28ad537b1b9c319d85d43bfa844a0eb8 +size 1587389 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041059/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041059/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..0385b8c45781c29ffc0039ce70c5eb0e7a36e0f7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041059/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:820a75bcf4cb1225fec8b738015307d7bda33b6cd8f23698dd68dc148436f8e1 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041059/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041059/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..cbd61d85450a6b36095b3ce62c4026cc0e92eef7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041059/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d84312fa7f9ec43d5fdadb743dedd6ac2a197d4e1a1a2e88947ca1eb1f2123ec +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..da2efe1eb07f2604d63b5bd6f52d551e7dbaf1ae --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7010189ea56a8fad59cf122b242fc82882c8702d9ad883ecbd27b97a2ee5312 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..143afa78636e6fc647f59f65766d11c5800f81f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80322042474fb800cbc7d48b3c5b52f020152150b820c8064659c937f40e3cb3 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..86069f912452adacb9005de66bbcec7ce86151f5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfd620f54287a48843b67876db3206d22cea03620336c01dc57c2f5574ce6057 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..d37cc06c2f18f3894297030d978ebe98a3d47a7d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041059/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:177ddea7d510a096cf12aaac25a8803646b9b47ffca151fffd10ba64d6e7c009 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/train_20260521_041100.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/train_20260521_041100.log new file mode 100644 index 0000000000000000000000000000000000000000..869990c4ffe7749b23157f05e1059afd07662b85 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041059/train_20260521_041100.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6beea3a329230750545c6a784cba27579bda16320d3cad45c8c35637106008b4 +size 105098 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..3183168b8dbeadae039fe5cc0d0499973dd6d2da --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041348" +dataname = "tabsyn_n11_tabsyn_n11_20260521_041348" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041348/tabsyn-n11-15215-20260521_041536.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..65bb6973f44f796b0b26dc60390d0b9700739c88 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041348" +dataname = "tabsyn_n11_tabsyn_n11_20260521_041348" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 50 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..64ce642d8884140fcadf94a7eea745ef79016ce9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7f8ccff68e9ce118201cfcf9df6212bde72021e51521b0a59b21b8114fc7412 +size 93 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/info.json new file mode 100644 index 0000000000000000000000000000000000000000..f51111dbcd20283a3359556ea02ad713c66622c2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:307c71c6b8e8b6c176390f9d682e4d4ff8c573e7f681f24a688e084d3b19244f +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/data/tabsyn_n11_tabsyn_n11_20260521_041348/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/gen_20260521_041536.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/gen_20260521_041536.log new file mode 100644 index 0000000000000000000000000000000000000000..c4f28f730cb2ea757e9e82eb586b722b3348c310 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/gen_20260521_041536.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01ff34dd291f3daee5a3ac516b0821184f34688c2fc50ccfa68f517a9aabfea4 +size 949 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..7f667c46365213e9731a66bc3569a57e80e87393 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb4af56026b26971cd8592bd7cef03399ab5b503ba0861fd8d9ead463bf70579 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..da079a7581d9dfe921915edc76029704b35c1fbd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:994c1658a7384b3e7ec0b56d050fd07320848a42bf870c88871cf39afaa667d7 +size 2597 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..6fd11ff239c166cd18252ea4630d26b50f8f9506 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bea1e0dfc3c48e3090c8d15cf8f20349df5d528c893d442d56a24a28a940e34c +size 882 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..a51bbd8051a99e95915f02d4f57e36ccc9efc39d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:459a226a3a82cff6453aaf9aa1d300baf8079f6bc79054744a95e1769e9a1b5a +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..06d0e984076588ef0e69ae9a9cfe8a48d90dabc2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29d59cca171a87184579c7412c687b27114bbd66913ef3a148861ecea67e4504 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/synthetic/tabsyn_n11_tabsyn_n11_20260521_041348/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/synthetic/tabsyn_n11_tabsyn_n11_20260521_041348/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/synthetic/tabsyn_n11_tabsyn_n11_20260521_041348/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/synthetic/tabsyn_n11_tabsyn_n11_20260521_041348/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/synthetic/tabsyn_n11_tabsyn_n11_20260521_041348/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/synthetic/tabsyn_n11_tabsyn_n11_20260521_041348/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn-n11-15215-20260521_041536.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn-n11-15215-20260521_041536.csv new file mode 100644 index 0000000000000000000000000000000000000000..cd218ffeadd1a88ba5347a08acccf6490bc8099b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn-n11-15215-20260521_041536.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be11c5af3ebc24bd727e5c66cb7dd037f57d7e4675504d9c4fff5d43bf83fa3b +size 1585804 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041348/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041348/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..a0afc1c42844f3ba67467adf90e385413f341ce5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041348/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:666fdac61352177e61e172aea02e6d8236b42895c91f59eea25ed921d4e28f63 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041348/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041348/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..08b5dfdff3c0588c5364244357b7291025c8418f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041348/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:406bca791d9fc701c797d2f24890319542cbe6d6f0634673c9952fe14c528595 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..a2363b03239f4340b79c49df3abf9e53c57828f1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c841f156db777065bcd6174fdf1d26c29769fdcbf20b536795d53277526eda7 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..b5b7921f68596caa10de9b95e51e1f7996625188 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82a2cbdfd8a1685d26edd6cd0ae9eea67bda707c2ca63b397edde861e3d48acc +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..7efe0faaf7acb0179022fe34f989ac89a20e823d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af3aceb4655d5cbe24012c24414dd76ea248f0525be2065e312ed95153a38699 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..ec3c7f78f4c5d2bc46d73d74aefe67a9597414a4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041348/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5721849ad2500ffc4242741b90dbc1ad3e18eff70b3a6126647f6aa66e8aa16 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/train_20260521_041349.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/train_20260521_041349.log new file mode 100644 index 0000000000000000000000000000000000000000..afbfd493bf4f43741ee7f25e177f641ffaccb2e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041348/train_20260521_041349.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2966685de6510c3184a3b7802661794521d4e9df0a3aa8a82535dac755654af +size 187378 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..c4287a4f17804ad99c3600671d8c4ad412baac73 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041652" +dataname = "tabsyn_n11_tabsyn_n11_20260521_041652" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041652/tabsyn-n11-15215-20260521_041742.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..59de066a849879d15178d9418dadac33bd834cb6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041652" +dataname = "tabsyn_n11_tabsyn_n11_20260521_041652" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 20 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..10cda754b21201483fa5d3ea513980f8d3645526 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:691dd05091cf19357400cda0420947d2248ce3ab5872cd188918d24ffc8296a1 +size 94 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/info.json new file mode 100644 index 0000000000000000000000000000000000000000..df76b287768d10f8506aa350f5e39711b6695e63 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e15db16b32ee6bafa588c53ffd4b90c04241260ae5f2cc2efd23bad830da486 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/data/tabsyn_n11_tabsyn_n11_20260521_041652/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/gen_20260521_041742.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/gen_20260521_041742.log new file mode 100644 index 0000000000000000000000000000000000000000..d5c1eb531a12b03b3fd114072ab6917d04e639fe --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/gen_20260521_041742.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d85aface0e115483e52e07ad214e618614494fdc0df3130ab09a7911c3ebab85 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c7bce79b1ed0d0db520b36233c9170a2db71cc58 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8df43704e8827552e3372118e591144a28d1e5da459dff127d8bd1fd0c28817 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..7136dda655560e4d646873c045ddd8309b79a420 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97f5a19275838bb1b73f01eeaf8a496bec398d3d70f9fece149e6045e07cbddc +size 2598 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..596219df327101baa91f89786e00d759bacc2926 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1f156860f4d1e47bccf1823117b2c41126963df236e6a7c4c742841766c4f37 +size 880 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..5e8d4d150ba4d942aabc1e6b9358ffdb3cca6c6b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9bb00364b8fa4bb0e5d22da8a117699ae8d172fc566e667ad811fb5df32de5a8 +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e6def01b988cbf4e6928dee8cb63e3ce4df0caaf --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3cc00be1c05cec37667f71de8fcc121c12521ccc86349e33ec9d86cb78b7ea5 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/synthetic/tabsyn_n11_tabsyn_n11_20260521_041652/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/synthetic/tabsyn_n11_tabsyn_n11_20260521_041652/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/synthetic/tabsyn_n11_tabsyn_n11_20260521_041652/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/synthetic/tabsyn_n11_tabsyn_n11_20260521_041652/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/synthetic/tabsyn_n11_tabsyn_n11_20260521_041652/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/synthetic/tabsyn_n11_tabsyn_n11_20260521_041652/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn-n11-15215-20260521_041742.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn-n11-15215-20260521_041742.csv new file mode 100644 index 0000000000000000000000000000000000000000..bb2bdf566e22ab695e5f2ac78ea2ea3ac959074f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn-n11-15215-20260521_041742.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8dd50f89da1330cececb1ae94630a47142f885a99b138b6164262e92316914f +size 1585879 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041652/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041652/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..8b5e8585d2753d0983cbc3b9b4dd8a93654ad5d7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041652/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3630d1537090a07c68cd1c7294f8c1f03caa293b5e74700917d72d2248c6db11 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041652/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041652/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..e4ee150c0a21bb6efbe70c99fd8f53e7989bf15e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041652/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f55903c217e5f86318f598fab2bb9c723cdf97d584f33adae7b619214b4c717a +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..0a969cb1db963bd0d33a6dcd37ae1bdbe64a6c31 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:344cabb8d111dbe0181d1aaa3522e099a55488ea1f6e13447fdc02bace97f055 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..a06315273f7eed7394b79de4e33a93226959175d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b62369b42b4be0fe9d8f6c83495d2cbe4d04b4cb21f15d2e55930165e9efdbff +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..e4dfe479fb9aec5eba7b159b91f61235a43e279b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29e90acafde4ce63e88398f5f5cfdf4afb221c7da4b239c403f6d5978a0bd8f9 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..a13a266252cffa9c78595ac8a401cf8bd08c9821 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041652/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf6ddd076b2aba46118b1fbabdf20a4e6859e837777654c83f508e938de0873a +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/train_20260521_041653.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/train_20260521_041653.log new file mode 100644 index 0000000000000000000000000000000000000000..eadf1c3d60086d28240111d34f06b03f04aa9429 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041652/train_20260521_041653.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03ea661bd6444c13bddb53afac18d12e926791f0fb4c163dda7fd0c6ce9a4d2a +size 51198 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..8da7ee0c44f03a8129fe40d57da74de21354808a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041859" +dataname = "tabsyn_n11_tabsyn_n11_20260521_041859" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041859/tabsyn-n11-15215-20260521_042643.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..102f30cbd949ef81ef1014ebacd4cf1ebcfce1dc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_041859" +dataname = "tabsyn_n11_tabsyn_n11_20260521_041859" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 30 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..f99237cec9832dfc253a7fe0913e54b0033acacd --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4a29a5ca1c67d835e62e8cad173eaea893d3d73528187307c0ddabe17a9985b +size 92 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/info.json new file mode 100644 index 0000000000000000000000000000000000000000..0c11e06c432d547a4065f34009b6ecad6fb08feb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93b4fa069dad3ba61d33c767cf959fa04afab353a6f56605946eddc151621964 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/data/tabsyn_n11_tabsyn_n11_20260521_041859/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/gen_20260521_042643.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/gen_20260521_042643.log new file mode 100644 index 0000000000000000000000000000000000000000..961a83d68ae0237865abbdf9e9c06d8c3f4ceaf7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/gen_20260521_042643.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0275c154e0a7a69bce65799f4eb9e37898a2cb79cb42b80baab3a529f8a509c +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..674b5dc467ed18e645ae2fa26a6d95f703f414b8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f3962f8bd56ea4960e44ef13b0d63c97cfcf6b56356238d16c17f4eee4a1b6c +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..427bcfb973106994ff2756b6d453609120d85fca --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae638a952dc351a0ccf37670155c49bf6d04a328e1024f8fbe8fad19e5ade2de +size 2593 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..bbf17ce9e68e51a33e896d05746b3d45bf38fd16 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2759b2724ab9557a8684bad791072ef71db69df7d486f3806442417b9a4528a0 +size 882 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..30c90989ebf7b4a06951b448b0a755afceb19834 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e0b9afa13d3043bdf9c4ca6b5eaabda3ce8f4ce8f59accdda70efe9bf556371 +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fdb3632b2b9cd09c269e877ae5049d052277c705 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd0344645c24bc4bc192e51029ca0b0bdbb8ba739f91ff7e7cdedead92df3876 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/synthetic/tabsyn_n11_tabsyn_n11_20260521_041859/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/synthetic/tabsyn_n11_tabsyn_n11_20260521_041859/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/synthetic/tabsyn_n11_tabsyn_n11_20260521_041859/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/synthetic/tabsyn_n11_tabsyn_n11_20260521_041859/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/synthetic/tabsyn_n11_tabsyn_n11_20260521_041859/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/synthetic/tabsyn_n11_tabsyn_n11_20260521_041859/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn-n11-15215-20260521_042643.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn-n11-15215-20260521_042643.csv new file mode 100644 index 0000000000000000000000000000000000000000..3b7cc41b39cb813a07411da5e9afe014f513a2ad --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn-n11-15215-20260521_042643.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc8b51d493b3bfb995137329c3f2fcfafa76228e2c76e973969df4accd9ef0ab +size 1587432 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041859/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041859/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..a198f5b33126ac1aa4556bee49c8db6940d9eb0d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041859/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9d8ba7ea2b0f7afce0ae839aa17b5518d842e2742dd49e757b1e9bed3d00097 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041859/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041859/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..e3db359b46e8a8e237b61dd697bd0d2f780d0fd9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_041859/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db2874518ed861675b548d1120c941aa283a047c83f26105c0b01c4ff22b2a11 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..14059588226c19356fdc8810fb5edfda5fda2f13 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1be40ff55c5e60c031fa0e798c6c416182bc7b1dd739c9fd00040567e87f10e +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..0f7908dc45853ec937174a1b864ab830f724c57a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bcb1fda71538526fdc89aa401f5fc6f3c26e6127fca66d0084ef7fae89dd359d +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..ccd6093a70167f4bff08dfcdf09c2a0a96d03e97 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d195f6e95f0f513486ce30d9a30369648afbd75d2f205dfed5371811a6c41413 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..c0101f48be112b59ca957b6dd2a93c6c32262f21 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_041859/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79e1e74d1897f8d32da5a01a3542efd0d2f424d681cefaae9eed7a52c6aee4c4 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/train_20260521_041900.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/train_20260521_041900.log new file mode 100644 index 0000000000000000000000000000000000000000..167af320fb657793bbc70c0994acd57ae7554d5a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_041859/train_20260521_041900.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c178f6abfb70317a5c2497208f0f511e6e8026d20320952c83069e52e02a7cf1 +size 424459 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..b2ed2e5a6efd5afd7236f8cd655e65f2a3071b02 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_042830" +dataname = "tabsyn_n11_tabsyn_n11_20260521_042830" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_042830/tabsyn-n11-15215-20260521_043330.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..e1b80033fb85a3ddb4d1e61e013902a02c64cb8b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_042830" +dataname = "tabsyn_n11_tabsyn_n11_20260521_042830" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 40 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..ae98784f06e66749135ca8458177146c820c8f08 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21609c39932c24ac06df452216adfd8cbd4648b4f3eb7a40099c4ce898666bd4 +size 93 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/info.json new file mode 100644 index 0000000000000000000000000000000000000000..a18252d3ad9eded8238038c631ae5fb6beeceb20 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e190c1ab16bc4024b9dd690a1f07c33900910fd904885346c5c3b1540fd7229 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/data/tabsyn_n11_tabsyn_n11_20260521_042830/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/gen_20260521_043330.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/gen_20260521_043330.log new file mode 100644 index 0000000000000000000000000000000000000000..cae89648f3a68e2d45b1a168d9494b7133c2d9a6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/gen_20260521_043330.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f1127205943792f85c536adcd9c68bf4705eeb10f790de7af54b2a7fb2cc1d0 +size 948 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..df7e62e3d924860786883fc81d499369166b441d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47cc832eaa4d626f0873d3dd6c0ec49c1b309f062ae7e0f5581e131e743ef8fd +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..ab92085e4bf9ec4efd9d826cc53e992074ad027f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53e2e5c6b571c4db1d4c9a6363864fbf4aba2f79888b9fc16f1e01dee19e0d5b +size 2594 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..41b3f5bc7c1226548a9feac050674e22355a19a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ade6c0eb7b8ea140c988aa4543c23cdd04fb0b23c3e919931a132748951a6786 +size 882 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..3b685ee5bd08b34671894a943c683160fa0e51d3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:303bb7aeed64e89a201a2bfa5128a3b3d2e5d04bffb64610e4508efd2ff3e224 +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..38b7a722707ed43cf03a2967173b5385ec807971 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b893ef713580da906d71d9562660d5d9ec8dbacfb405287b19774558e94e566 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/synthetic/tabsyn_n11_tabsyn_n11_20260521_042830/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/synthetic/tabsyn_n11_tabsyn_n11_20260521_042830/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/synthetic/tabsyn_n11_tabsyn_n11_20260521_042830/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/synthetic/tabsyn_n11_tabsyn_n11_20260521_042830/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/synthetic/tabsyn_n11_tabsyn_n11_20260521_042830/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/synthetic/tabsyn_n11_tabsyn_n11_20260521_042830/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn-n11-15215-20260521_043330.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn-n11-15215-20260521_043330.csv new file mode 100644 index 0000000000000000000000000000000000000000..ad96bfb850a94b93f4ef7959af194390b351d5d4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn-n11-15215-20260521_043330.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8dfe281726183b2c43898b244bfeb707e9cfea355b01326052d58051cae207b5 +size 1585635 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_042830/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_042830/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..cdbfaba7845a97988cc370e7ec3c8b4d3770c4e1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_042830/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0def787e568d4c295ffc4f5c2ffbc79a35a872c46fa2d8384069daaa7684d571 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_042830/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_042830/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..317ce70a77d0e85fef4fe8d1a92e56d23ba290d6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_042830/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:780c338069ac015b562b8e348f409bcee0e5d81d042c429abf7793d88e3b5536 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..8b8919c9c57549ee2f7ae874700f6b40bc3a4162 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edeb982b8f89e87093b7a58c1197c60e2224e5872ffb52b64fdb831074a62a09 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..1aa9314ffd87352e5e642ae70141e52e936d0f3d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47719640efbd5f98f029049ac4245e6c60178b361830d8a1e613c12c22a41674 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..f6299a8750b91d35434d818da46f6edba2ac953d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02839e30b17626750ff1d06e2145b2b8769b1be3f94507004ffdde0280ad8f66 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..8fbbbecf11f90f2a5ef35224094904bfc42db546 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_042830/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eb586555b916d80ca8d0b1e580193267a0b6401ad7faf1f879f7051f046ab05 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/train_20260521_042830.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/train_20260521_042830.log new file mode 100644 index 0000000000000000000000000000000000000000..d85d9bf1ee4869f35602a3fc0c0ad3fde5245f85 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_042830/train_20260521_042830.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2027cfd80638bf679e815d6453e5174e2070c1c773febede8c47037d4e493a2a +size 229670 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..3e7d4944a47245214b538dc1a500139e67182d45 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_043448" +dataname = "tabsyn_n11_tabsyn_n11_20260521_043448" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_043448/tabsyn-n11-15215-20260521_043513.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..093b78f269843af959c860bc312ef45363dab607 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_043448" +dataname = "tabsyn_n11_tabsyn_n11_20260521_043448" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 5 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..c15259d76a06098db2fa20cad209b3d4337eb9c4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f10360316347c1da81b575e0c28e35910ffc4df1608f4762c88bb4f73bcfe2f +size 92 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/info.json new file mode 100644 index 0000000000000000000000000000000000000000..4c9686863404bf1aa3928d1ac0d764cf6974827c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c48ba5cb37d7b1895198e1f910f583a8d514b9a88b4841f8c67f513991606f07 +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/data/tabsyn_n11_tabsyn_n11_20260521_043448/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/gen_20260521_043513.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/gen_20260521_043513.log new file mode 100644 index 0000000000000000000000000000000000000000..97560a334245cb1621f363923a0622ad2959532e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/gen_20260521_043513.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06f1476929467828e28828829fbb5ae896bddf9752ac92a87761f351f9f74886 +size 949 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..830d4e964890a4cf087f154fd70e5755cc88b7c4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5dda96bda4acf59fdb3424c4fcaef0a14a42d5033236497b0d04c3557c6e24a4 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..62f57b40f1da582485f683260161beca2f60a23f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2e36f915093c416ed41afcc059c692def89b6d8509dcd6618d0edcb5b596f2d +size 2594 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..0330750ccd8d90bdb4303dd6fa424b6ffbb55857 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19f3f166d2cf02aa3fa39168634fbbc172f8178231790b34e7a5874f43d8eca0 +size 881 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..afc35513a9d473838f715963a18db8b295726b1b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b8f000fe13d2030daa9acc629a24dba4753713372659f8536981544100bf604 +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..325aa528771c0ffe8387150889d9264ccbfd2fe0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9438ee8dac3e877b465b4cf89d6b71d7750c9ae2d618d37c61a640a7270dbc4a +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/synthetic/tabsyn_n11_tabsyn_n11_20260521_043448/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/synthetic/tabsyn_n11_tabsyn_n11_20260521_043448/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/synthetic/tabsyn_n11_tabsyn_n11_20260521_043448/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/synthetic/tabsyn_n11_tabsyn_n11_20260521_043448/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/synthetic/tabsyn_n11_tabsyn_n11_20260521_043448/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/synthetic/tabsyn_n11_tabsyn_n11_20260521_043448/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn-n11-15215-20260521_043513.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn-n11-15215-20260521_043513.csv new file mode 100644 index 0000000000000000000000000000000000000000..9765a0e041d8acb3825c442adab10bffd0794ddc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn-n11-15215-20260521_043513.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1efde555fc0a485f2e3e8dfda2bcdd49722e14d712afd15bd82f7bb3ca395bac +size 1583326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043448/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043448/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..a12a7e808761cdc8935482387c7c894b708dfc5b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043448/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f200136c27839061b579708e726e939d7b8a360aa86fb5cc06e48d1072bc279d +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043448/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043448/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..788dbcbd2cfa89c26b9272ded34f85cdb8ef48b2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043448/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:953b908390e472a3ec5a4894d5f54e89233e166aa17034ec427e01a895fc3d7f +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..8f572f456afabf53f1f784fbcbc06c0bc911f585 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ae17bad824b643de25e722dd22b3a96d525598e8f9de9021524a1df5c811593 +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..91a8310f358e2e208a5028292d97027d18661c6c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b07742d0a2a7a267ce3c6811133874a8ac6fffbbd085aa3680de2af52b7bcf38 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..81222f0d2a15ad51a19667119d37719dc0c01eb4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65ef5aa698945ba5ea2d0a15ac97193376ed4a085dd78406043fea2bdd702bb9 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..ed2782813079931154505dc7bccda113d0742f6e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043448/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1138319bb9754af20a9c2735afcad43e1b5f5010fb126e36ccfebf8dc6433c27 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/train_20260521_043448.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/train_20260521_043448.log new file mode 100644 index 0000000000000000000000000000000000000000..d44af981c2bc6e7a11bb6269e592188d27db2130 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043448/train_20260521_043448.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8aae0445c294316b9f7c034a7a6aea6c7bda4090895ad00c01f2850db4650d0 +size 38553 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..2280a9220fb33482594a27b482a5dde47d2d50a1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_043630" +dataname = "tabsyn_n11_tabsyn_n11_20260521_043630" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_043630/tabsyn-n11-15215-20260521_043722.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..1859c01ddd5e28eb245d3a67fc2ec5cb1b9006da --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_043630" +dataname = "tabsyn_n11_tabsyn_n11_20260521_043630" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 5 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..4ffbe8ba90f971f129fd4e2c3987e3da541459c0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8760697975d02e3d7cec4a46208422b5df5cdf68fd825501d726b711988ee56 +size 91 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/info.json new file mode 100644 index 0000000000000000000000000000000000000000..0575a41b4a70ca8863cf99b8cec7625d6bcf3e26 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1a52db8222d2c3774711888ccd71d5c0c7dd740264608614248aa6b6ac0074e +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/data/tabsyn_n11_tabsyn_n11_20260521_043630/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/gen_20260521_043722.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/gen_20260521_043722.log new file mode 100644 index 0000000000000000000000000000000000000000..a09a19c21bc82c7314707adb8b3032fbf039ddbc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/gen_20260521_043722.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7e2e8d76c1d8600bb30913861f0e4d9785e8fc6aadedb436a9bcc63453231e5 +size 950 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..db06ef7276363ee79e1ac47df50e9a460e118072 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2850f2e37bbcd09fe19cf6db004b53b0e375bf794652ee441efe266c58ca295 +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..c736bdf0583d0a08c7e8b042236ca23449cbce1b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a82a009db6fcc3b7fec6ee3e26868551b9db92ff33f65159e344076fd6ea1a9 +size 2590 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..58d338bc0769fe16ea904c169f6537c9ff2fb15c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f643dd71cd574e9baf8fb0d48f81a50e7e958bad7daa481ac31ec314d5fb3b12 +size 881 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..66a94fe3e134e339fb9f1108b516576b38102ee6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a418f50ff4354d9634d167a905fa34e17bb29e7c24f952de904bef4a1a30e2f +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9b61ecf64e59a96eccb96f6549a6d7d2ac5f7fd9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8b1eacff798f49fc80c9b84b660d0110376ab8d40a407bbade154bfbd7b20d2 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/synthetic/tabsyn_n11_tabsyn_n11_20260521_043630/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/synthetic/tabsyn_n11_tabsyn_n11_20260521_043630/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/synthetic/tabsyn_n11_tabsyn_n11_20260521_043630/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/synthetic/tabsyn_n11_tabsyn_n11_20260521_043630/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/synthetic/tabsyn_n11_tabsyn_n11_20260521_043630/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/synthetic/tabsyn_n11_tabsyn_n11_20260521_043630/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn-n11-15215-20260521_043722.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn-n11-15215-20260521_043722.csv new file mode 100644 index 0000000000000000000000000000000000000000..54632ffce51ce716f90cc581cea0e33f2699a613 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn-n11-15215-20260521_043722.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:409ab33afd43a8ebc9d18c132dead57da492c343e00d855db303b35d96a867c9 +size 1584079 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043630/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043630/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..e5c179013ad6e0df28ed2128e200f5f9f95a1e0d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043630/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d4afa82a0b46d8bd098fc682047c47f58cff937b48f04bf9cb53ca79833ec95 +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043630/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043630/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..0b6da9fbac27d3b1df718a0c7b164e1c5cf9188a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043630/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28613a14c6e72ac03d70b605306bd9dd68ee81253da236ff5183858553c85faa +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..fd8c9a1b96c3822d3f3e0c191bfff7f22e10988a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:444fa1346af492c6b386fa7e802747551ea43756752b944ad0d1f6015ec326af +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..97fcc7731e13fe7ebf01c36f8e76608e3178b8d3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c903b599e0f668eed2aacad75ff83b9b4f5ef5656881abe39223666a2304d8d6 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..91c93a86f4c5568d4d1db0d58984acf1cadb6229 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e426d2835844ab266132f4e3214e9310064a3db7b9cfcd9dff2bedf80bc22370 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..3c55512285647b1b1ec686afd98dc76f114db432 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043630/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:473f70fa6da5ae034e16eb6367e0546afef82b1e45b60ce85d4e460cf9559149 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/train_20260521_043631.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/train_20260521_043631.log new file mode 100644 index 0000000000000000000000000000000000000000..66920e9f851c80a89ad98daeea9aff8291671a12 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043630/train_20260521_043631.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e08514f071609f0e9331937d06a696848e89799a9b6fa894b9cf1188ac990842 +size 38887 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/_tabsyn_sample.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/_tabsyn_sample.py new file mode 100644 index 0000000000000000000000000000000000000000..1511dc02b1ed9e50bc70f9b67eff9e8077b4223d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/_tabsyn_sample.py @@ -0,0 +1,43 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_043909" +dataname = "tabsyn_n11_tabsyn_n11_20260521_043909" +output_csv = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_043909/tabsyn-n11-15215-20260521_044102.csv" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Ensure data symlink exists +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +print(f"[TabSyn] Sampling 15215 rows") +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "sample", + "--method", "tabsyn", + "--gpu", "0", + "--save_path", output_csv], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + sys.exit(ret.returncode) +print(f"[TabSyn] Saved -> {output_csv}") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/_tabsyn_train.py b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/_tabsyn_train.py new file mode 100644 index 0000000000000000000000000000000000000000..a828c087729674bc7cacd59778f4f69b6ea59396 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/_tabsyn_train.py @@ -0,0 +1,69 @@ +import os, sys, subprocess + +work_dir = "/work/output-Benchmark-trainonly-v1/n11/tabsyn/tabsyn-n11-20260521_043909" +dataname = "tabsyn_n11_tabsyn_n11_20260521_043909" +tabsyn_root = "/workspace/tabsyn" + +assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}" + +old = os.environ.get("PYTHONPATH", "") +os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "") +sys.path.insert(0, tabsyn_root) + +os.chdir(tabsyn_root) + +# Symlink data dir into TabSyn data/ +data_link = os.path.join(tabsyn_root, "data", dataname) +data_src = os.path.join(work_dir, "data", dataname) +os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True) +if os.path.exists(data_link): + os.remove(data_link) +os.symlink(data_src, data_link) + +env = os.environ.copy() +env.setdefault("TABSYN_RESUME", "0") +env.setdefault("TABSYN_VAE_BATCH_SIZE", "32") +env.setdefault("TABSYN_VAE_NUM_WORKERS", "0") +env.setdefault("TABSYN_VAE_EVAL_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_INFER_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +env.setdefault("TABSYN_VAE_ENCODE_BATCH_SIZE", env["TABSYN_VAE_BATCH_SIZE"]) +# Safer defaults for wide tables on Docker: reduce shared-memory pressure in diffusion DataLoader. +env.setdefault("TABSYN_DIFFUSION_NUM_WORKERS", "0") +_te = 40 +if _te is not None: + env["TABSYN_VAE_EPOCHS"] = str(_te) + env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2)) + +# Data preprocessing is done on the host side (_prepare_data_dir) +# which creates .npy files, train/test CSVs, and info.json + +# Step 1: Train VAE (produces latent embeddings) +print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "vae", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] VAE training failed") + sys.exit(ret.returncode) + +# Step 2: Train diffusion model on latent space +print(f"[TabSyn] Step 2/2: Training diffusion model") +ret = subprocess.run( + [sys.executable, "main.py", + "--dataname", dataname, + "--mode", "train", + "--method", "tabsyn", + "--gpu", "0"], + cwd=tabsyn_root, + env=env +) +if ret.returncode != 0: + print("[TabSyn] Diffusion training failed") + sys.exit(ret.returncode) +print("[TabSyn] Training complete (VAE + Diffusion)") diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/bo_combo.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..00e648a7e28fc9c4d6f127b0e98f47ee132224a5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c028375b3891d05153998f62873572c3af2d265dc24a490596272f1e6d1836cc +size 93 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_cat_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_cat_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_cat_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_cat_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_cat_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..65f01f4aaf9b5e6ac281ddfd05f178b82dc24ba9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_cat_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6735c7290e805ffeb57d96cdb652930cd19766131eb333e144f301ca5c5d90 +size 128 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_num_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_num_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_num_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_num_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_num_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..cda65645f7509c4b59dba4a620ca6db54dcdbe84 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/X_num_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013e32900eef106da63d972b40b61296d91c04d1ce8196fd0cb8d54081f58c59 +size 608728 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/info.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/info.json new file mode 100644 index 0000000000000000000000000000000000000000..7e48e8e28dc3a5c26b70428a25177ad0d47fc453 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/info.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:298526def2a15d5e250349d56bba5dbcca8c7c3ef74ebf35f74fa5bb6539430d +size 2286 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/y_test.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/y_test.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/y_test.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/y_train.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/y_train.npy new file mode 100644 index 0000000000000000000000000000000000000000..b9daa4049b7c19eb916f95443c35c7cfbf34726e --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/data/tabsyn_n11_tabsyn_n11_20260521_043909/y_train.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed032a868b2263b0d8eb889ebf4ea480e38bcbbc5c51dedc15645d235005d2f9 +size 121848 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/gen_20260521_044102.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/gen_20260521_044102.log new file mode 100644 index 0000000000000000000000000000000000000000..b2c400c5565ffdc191b0b111e50a6fa9118ab78d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/gen_20260521_044102.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09390148a8d5fadbaaeb4062e65e9c86f249f0aeaa6e4e37440bd58d6c8c1b37 +size 949 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/input_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ea2be41f2d4e3c19baf33a693e51d842c71b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5079ada73320e6a27fe472104e990fdf996a6bbccaa9cfdf3c72c6f00bafe2e +size 1359 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..1b828a8b99499f32f3363c7bc9694c6f1b6baa43 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f95f709d78620609a36f71a8738b971afc9c3b841073fc0682d294f2de00c34a +size 6089 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/run_config.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..18a0d1e7e09e01c1657997e12dab74e5f3e34999 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:329228da08c3a45e692f7f66fc3ec2c0c636498ab06c017fe893a959a0294302 +size 2597 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/runtime_result.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..d9d9bf20dcd1fb26a8ddb5acb600d01383d77eeb --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24ac75df7d6e10a7bda34144e255144657e46c1a2e055cbfc66fcef98b61ea7e +size 882 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/train.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/val.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/tabsyn/adapter_report.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/tabsyn/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..50e09dc5389911918757e0b3d4a2a89b0665d7ed --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/tabsyn/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f0a084e78e6e2c1a5fbea378f2050f4859c48e4eabdabdc3237a8773e11e864 +size 320 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/tabsyn/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/tabsyn/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/tabsyn/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/tabsyn/model_input_manifest.json b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/tabsyn/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..789307cb95ba5d5a50da61c0a3e77ee601d9f7a7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/staged/tabsyn/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bebf24cdf77e1043ee4b3a73ffae6f4b937a92436c469aa8d1340985e12be363 +size 6285 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/synthetic/tabsyn_n11_tabsyn_n11_20260521_043909/real.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/synthetic/tabsyn_n11_tabsyn_n11_20260521_043909/real.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/synthetic/tabsyn_n11_tabsyn_n11_20260521_043909/real.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/synthetic/tabsyn_n11_tabsyn_n11_20260521_043909/test.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/synthetic/tabsyn_n11_tabsyn_n11_20260521_043909/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..adb2da2a9734fb5b3bb2e0fa1cd5b5815b1416c8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/synthetic/tabsyn_n11_tabsyn_n11_20260521_043909/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976bc9e44218596dd6aaa5480a6b4a586c49f9d705a939a7c3089af9bd29cf68 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn-n11-15215-20260521_044102.csv b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn-n11-15215-20260521_044102.csv new file mode 100644 index 0000000000000000000000000000000000000000..e04532aa68bc57f4b92fd0af4f0b7e7027927d7c --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn-n11-15215-20260521_044102.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:291192f27721c99fd59f18012521e660b5d574d6c44d271aabc8d6bfe1b6816b +size 1586206 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043909/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043909/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..ec3a2dc9c55835485ee6e74af99448fbc0bedd9d --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043909/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c25ea9d9f2e3fdf1f550626e95f725ac71d7db9cd4a2fd64e346e50bd6edd33f +size 42341841 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043909/model_0.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043909/model_0.pt new file mode 100644 index 0000000000000000000000000000000000000000..d9787af7fdd88493a5181a9732edd1e74e1472b6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/diffusion/tabsyn_n11_tabsyn_n11_20260521_043909/model_0.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c72ba017aef7e3256063dabbb553ac13ba5cb73d746fb4554c4fe850373b3bb0 +size 42342137 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/decoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/decoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..4eeeb214a2647f184d86df1a4e2c978b1db2d28f --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/decoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d56e6206588d564e87e059205b9cafc48b05eeabde43c14bc7c4ffd2b5f2b7c +size 21888 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/encoder.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/encoder.pt new file mode 100644 index 0000000000000000000000000000000000000000..2b64e85591e006586d2aad383605c6dc078690d5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/encoder.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de8786b6d832627f2d97aeec67fd7891533f61f43a34921d0cd22476cb44e536 +size 22525 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/model.pt b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..2c3418e897b6b73f4c0ac37cc481353203535c43 --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc770dcdd03c41956114a9d1a5cd7b1178d14118e74c58d2967b9522b21f2e21 +size 62706 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/train_z.npy b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/train_z.npy new file mode 100644 index 0000000000000000000000000000000000000000..ae7ed4da1d320c3ea8c5a2c21e6e54b2a81915dc --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/tabsyn_ckpt/vae/tabsyn_n11_tabsyn_n11_20260521_043909/train_z.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fe546708cae40313e12029f792e8b85f81695f2c10dea276944b070d0304940 +size 2921408 diff --git a/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/train_20260521_043909.log b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/train_20260521_043909.log new file mode 100644 index 0000000000000000000000000000000000000000..4e109dd59ef9fa4e19da3f535b1b662b90560e2a --- /dev/null +++ b/Hyper ParameterTuning/n11/tabsyn/runs/tabsyn-n11-20260521_043909/train_20260521_043909.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9bb949fac9e927f11e585718006d6a158b359d914f7190ac110ae1b4fb24130c +size 483639 diff --git a/Hyper ParameterTuning/n11/tvae/_rf_tuning/final_recommendation.json b/Hyper ParameterTuning/n11/tvae/_rf_tuning/final_recommendation.json new file mode 100644 index 0000000000000000000000000000000000000000..3448ea0cf50e0f7b8bbd5f0dc4224061e4ccf6df --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/_rf_tuning/final_recommendation.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a67764c8b80d7d32591899d276f1d7d51a030aa47f99125455e20ebee5b6bf7b +size 3475 diff --git a/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase1_manifest.json b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase1_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c4e5ae1e644af7f2511ad514008b07686e7ef959 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase1_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:127d72d489be3add392a373e3a1581e3f0098b946cc8e4fb66cac337e07dbf2b +size 4487 diff --git a/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase1_scores.csv b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase1_scores.csv new file mode 100644 index 0000000000000000000000000000000000000000..a5c92cd8716e463c03e7b5be372217a388d965fa --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase1_scores.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:605c8cb17b26461b58df9080187ade3deabe51afc134c2faf4a571e2e14ec9b1 +size 3642 diff --git a/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase1_summary.json b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase1_summary.json new file mode 100644 index 0000000000000000000000000000000000000000..b39610ce8f6a641268fe32100ac9632e3a2cb759 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase1_summary.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:013d1e03c20462992e4e76834224296af4ffa4e7316c5874ac2c948fb9530073 +size 571 diff --git a/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_manifest.json b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b0a698d42340fb300d9fcfd33ecab991b91289a7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22b232b02c906880ccb4e67b883291050b8a99c6fb4d1e2f71bb6dc2284dd62b +size 3774 diff --git a/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_report.csv b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_report.csv new file mode 100644 index 0000000000000000000000000000000000000000..44e3a0ec473a54bac2a7d95a212298b5c657b8e1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_report.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc04f4f1a7c657c5d8c6873c864d2848356e551aacc8d265883bfce1c8b178af +size 1137 diff --git a/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_report.json b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_report.json new file mode 100644 index 0000000000000000000000000000000000000000..7bf9a05395c5c98330a933f905172952fe449825 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4761978d91a181f6050ec8c3e4725e138476db680d3305a8e3b353c89905076d +size 4132 diff --git a/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_scores.csv b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_scores.csv new file mode 100644 index 0000000000000000000000000000000000000000..487d7a207fd41c0149298d5c2ff75a0707421509 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/_rf_tuning/phase2_scores.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a25175057271f0b40199aa14677a8d98400b856f935756016aa45edf0836fb43 +size 5222 diff --git a/Hyper ParameterTuning/n11/tvae/_rf_tuning/sample_manifest.json b/Hyper ParameterTuning/n11/tvae/_rf_tuning/sample_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4f2921b22a4c1b0af0405d1f80444b5b21ecfe96 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/_rf_tuning/sample_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecaf6990d8e247a857755809c23832df485db3060ff2ff53b23d065c0fa033a8 +size 631 diff --git a/Hyper ParameterTuning/n11/tvae/_rf_tuning/search_space.json b/Hyper ParameterTuning/n11/tvae/_rf_tuning/search_space.json new file mode 100644 index 0000000000000000000000000000000000000000..37c6e357d3fbdb9ee4d4c26daac73e09743ca32c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/_rf_tuning/search_space.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba831e2e580e183329acbfa611025c5a1197625b9ac7436108b41e72ceeb8295 +size 1000 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..6daafe652181d2944480e2aaeff3cf902b7abc20 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_084649/models_100epochs/tvae_100epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_084649/tvae-n11-15215-20260519_085119.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_084649/tvae-n11-15215-20260519_085119.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..f0dbd9a71dec7cb518b2ff2eed3325131c9b4d41 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_084649/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_084649/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_084649/models_100epochs/tvae_100epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 100)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..095648ef8bea159ba6121aec51cfa4966e352174 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f3c8bdb40c7df909f6c8d94d455c6b4125f5f311216d93bc86b90ef8665481e +size 286 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/gen_20260519_085120.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/gen_20260519_085120.log new file mode 100644 index 0000000000000000000000000000000000000000..43a5d5eaceb824012e646a3288e668687c998cd1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/gen_20260519_085120.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0c0470c5f177da0353ca37510170143495191a8dadc7859a96096c672240ce6 +size 408 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/models_100epochs/train_20260519_084649.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/models_100epochs/train_20260519_084649.log new file mode 100644 index 0000000000000000000000000000000000000000..cd46df2eb3d731b52e04b5269ea6d4f01dc998d8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/models_100epochs/train_20260519_084649.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1824c281c58f7846732dc76c901f4dda977364dca0d016cb1a11a9a0210b32a +size 682 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/models_100epochs/tvae_100epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/models_100epochs/tvae_100epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..0042243651c60603d23863939a1e1c46304764cc --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/models_100epochs/tvae_100epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eebd23a7c0969b00f1dd1c70652b0dc9deea4d8c76a41be1830b9db8282430f2 +size 1569202 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..25e60ba1b63dede2051fb922fff32c76b7a8fec7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f11081680d8f39a584bae847d33b63e70d980fadafba6cf655248da588781a4f +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..e1fee7d55be6cd9a432b6e29f90260f4c4783957 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:340b58ba76607324f167eb86563b514aa1b52d89034224de7671a848f6ddd00c +size 2455 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..c3520d48528a7eec8a298ea19be0eaf8a68a6245 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eaafa912e7ea0547dbe57b1be449008902bbe583e01b18876270b6736a45b48 +size 903 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..15fddb154a2397cc7459f8a2fe9fceebd7e0d4bc --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a1c8dd7b71524a3f849d6d08fa5fc6baaca00618e39f64df758d067407fe76e +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d2d37258ab21f0795a5dcae684008d2941734316 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b820de5f32fdd9ebf09fbdf39582d2645614030c94c02ec13fd79db4a0790de +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/tvae-n11-15215-20260519_085119.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/tvae-n11-15215-20260519_085119.csv new file mode 100644 index 0000000000000000000000000000000000000000..9fd6c2ec1568a8b3cf2ac1eb630db3d6d7a79c53 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/tvae-n11-15215-20260519_085119.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a31ae099e75d41169fddc87424c0d195fee79464fff4466494d10824c9dcd028 +size 2883069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_084649/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..a4d888f45e1dff7a05e45aea9b93180616321143 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085217/models_100epochs/tvae_100epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085217/tvae-n11-15215-20260519_085635.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085217/tvae-n11-15215-20260519_085635.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..623923a6d5aaeb9915eabee9cfca39cfe7d3ff3d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085217/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085217/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085217/models_100epochs/tvae_100epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 100)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..7066ab0cb1ca1f60db12571ac29b16ad930901b1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:045228d191f96da0db83bdfd94abeb9cc7e9848abd8262fbcf6ad9ab6cabde94 +size 268 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/gen_20260519_085635.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/gen_20260519_085635.log new file mode 100644 index 0000000000000000000000000000000000000000..f08d7ea73dd63b872c57a9fca3b73fb4159c2970 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/gen_20260519_085635.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0976434ecde5e95a0728bf112c16a6c348baf9bde51a69e38b402887312fd672 +size 408 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/models_100epochs/train_20260519_085217.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/models_100epochs/train_20260519_085217.log new file mode 100644 index 0000000000000000000000000000000000000000..83a4bb7346f247aa29ebfde6ddb3dd8817b363df --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/models_100epochs/train_20260519_085217.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3213b209696e772ea4d849a4670eccae0d5a24b1bc5a075ac7e16333066e811f +size 672 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/models_100epochs/tvae_100epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/models_100epochs/tvae_100epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..f80173e485cb39ae1f982fef50b5b5069eb9ce74 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/models_100epochs/tvae_100epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e6517ababe3571454915fed0dca63690e1e3ccc8dca5683eaf62a9fb51e1ab4 +size 2207148 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d12571584527c046aaf05a8ceade475fc67ac62f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54461e896e125cc0e2246ad623ba0a6ceb760ddc8aa340e69f35ee36c059063a +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..708311106f59e74fdd21d5b652b28ae5a75d8188 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:354b25dacd7f92df07adbd7c7f381c98bdbab732a5980b6618de5cd754263d6d +size 2447 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..3a60125731352c8f1febcbfe030a34808cf3391b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:406306544e75d5844d4ce202893d30a1196b10a2d130842e002f6fa221902eea +size 903 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..44b874664ed9349f552079dd39e1a93f4cc901ca --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9c061bf4d7b401373e75aeeb9eae3fa55288f833fd28a1f44b81016cee6f09b +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..3774994a8504b86d2ec90373d57bd9a7de04ab24 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a993cf8102c5821331dbf976a633113bf22c4c92f13ad4681e888f13c494acb +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/tvae-n11-15215-20260519_085635.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/tvae-n11-15215-20260519_085635.csv new file mode 100644 index 0000000000000000000000000000000000000000..3fc59281b3e26aa9e92035f5f2c1b9ba1521069d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/tvae-n11-15215-20260519_085635.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a73dab0dd67e4d8eac1d8067ececcc4fb3e354b55282731efac914b5ed8daf4 +size 2883163 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085217/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..748c7884015bf329449dbffed4fafdbec53d8c86 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085732/models_50epochs/tvae_50epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085732/tvae-n11-15215-20260519_090138.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085732/tvae-n11-15215-20260519_090138.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..6f1a088af46467887d369ce829bf062ca4aa1afd --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085732/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085732/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_085732/models_50epochs/tvae_50epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 50)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..51f1bfa2880df4c194923604d0fdcd8f490a657f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc4251b3396c8713da36fa84b6fa3c221d23d55df3cbaedca8706ee559d7f987 +size 246 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/gen_20260519_090138.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/gen_20260519_090138.log new file mode 100644 index 0000000000000000000000000000000000000000..3cb4a2735b210124e8adabf10dc81b2683d9f155 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/gen_20260519_090138.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7073f58a22421b8159bb46c9cb40d5e016dc5c8553369cf626e5d17caedda9bf +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/models_50epochs/train_20260519_085733.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/models_50epochs/train_20260519_085733.log new file mode 100644 index 0000000000000000000000000000000000000000..337a36df343c729ca1faae490ddb9e4e08a3e0ca --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/models_50epochs/train_20260519_085733.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0daea594376c56fd0220d051e436714c1118cbf994f80eeca931e17809cef90a +size 658 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/models_50epochs/tvae_50epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/models_50epochs/tvae_50epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..ab443cc625a1bd62136f8102b1019e100edb09b2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/models_50epochs/tvae_50epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:def51b98c31fd4cebb6815a3b0a960de5c01c7086491fd300cb0dccc25bf261f +size 1029275 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8dbf9a1fcaecb560cae176c393c2794993a37a21 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c145cb231a5ffda2ca1eab61f1eca7038b3c551e4dad2727f1815ea3b086cdae +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..795a7eb174236f4a5f1e65a908ba19c501637b99 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:818c613c31c88a1a57784a0e8c6600b12cd7488e7941f78f530fbc5c2e337ed3 +size 2432 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..8a0777aaeadc680d01f88dc8f794e9919363e993 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26d15689b2488eac417e28f75f138806c0720bb9f11e468a3bce24a16eb3f0c1 +size 901 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..92e1ec1b3fa3e45f05d8ad12326a73919bd2c7e8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0269af872b93f91c30ddc52763c29e9f11927d60b28514853c8d26c6e86449d5 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c4e80e6c22ff020cc0dc3d931609b13b54878500 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a7ebd3f1c699bbcf9ca6c802e9f9ba03a131ba50394baba40db6f991a5e6b70 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/tvae-n11-15215-20260519_090138.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/tvae-n11-15215-20260519_090138.csv new file mode 100644 index 0000000000000000000000000000000000000000..494a41f5636de49be93df913d54a3cb94a9b8950 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/tvae-n11-15215-20260519_090138.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02ebc057a9c8c35c3275e94f2da5860ba00756aa8e230ac4146f420641d5b914 +size 2883960 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_085732/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..de6c36353f3e5235bed19f6dee4bf69dd1aa18c0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_090236/models_250epochs/tvae_250epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_090236/tvae-n11-15215-20260519_094127.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_090236/tvae-n11-15215-20260519_094127.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ec28e62faad2b7a7c13e004871d254b18fe06110 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_090236/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_090236/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_090236/models_250epochs/tvae_250epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 250)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..effcffc3695241e5e8911ba7a2469c39aa871b51 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50aaf9fc0c77b21bcc247d0ee29285326213bf10ee2e2e6af0ddf77af44eb8db +size 284 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/gen_20260519_094127.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/gen_20260519_094127.log new file mode 100644 index 0000000000000000000000000000000000000000..e5abaa0d0df62497247522d2f76bd4dc73aedd38 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/gen_20260519_094127.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:699436cd8f0c3868299d2f2ea5115e38e01a12f0c0bf9981bcdb3ae9bb0e46b3 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/models_250epochs/train_20260519_090237.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/models_250epochs/train_20260519_090237.log new file mode 100644 index 0000000000000000000000000000000000000000..db7d5d3bb24b089432cf7e36090f6fa74df6d729 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/models_250epochs/train_20260519_090237.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a1352d29d4944dc206e73e7e8fce0e686329a6e6f9343c5dcd5ba0780209dca +size 681 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/models_250epochs/tvae_250epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/models_250epochs/tvae_250epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..42a94619345e2db453753763195c4953f60d24f9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/models_250epochs/tvae_250epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22776d01a0923e634c810fd4a9539f1c2a29a2d96f208a75323f735e1ba4f0e8 +size 5928050 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..303b1adfd474ce17d35afa82304c89bc8e2c7797 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dda4dd39e6dc4174c8bfd112d54edfda937629a0991c47d5b6b9b14f19a67c1c +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..6b31b8c624bb6358dc5193e28f1e703f87b3053a --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a4ac3f88dead8599978503533527f48c6524ea6be93d520bc003e3abe280d7a +size 2453 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..79691e0d506013c739e64ca161ff149910a0070d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb8ed0179218fa8d6ff1983bc682c34c26883c0ac09d8f0b1aea6bd99be5dad8 +size 904 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..d1086825bc2910ebb88753ee580762a1addd2d29 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66d1d5ff63ec0cb8c4a94a629c60546698f0eb44dc3b89b52aff6b2ddbd06b11 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b42f24d88840e7c5903dda373b24041936354451 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:526d302a60482c8c26a829d8ba982e865c4fb7a9fdba0b12fd019631ae320f72 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/tvae-n11-15215-20260519_094127.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/tvae-n11-15215-20260519_094127.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0447139f0b9a78e4cfb4735b500f2948d9e2811 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/tvae-n11-15215-20260519_094127.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5906dbe2faa8d60ad40d87f81cedec838449d01d4f3255508387de7b330169f +size 2885017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_090236/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..d39bf5555299a7b2b62e6ef8edfd536ff7adad48 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094227/models_300epochs/tvae_300epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094227/tvae-n11-15215-20260519_094851.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094227/tvae-n11-15215-20260519_094851.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..63517389e6129898d0c4bd50ede91f85b0274a8d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094227/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094227/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094227/models_300epochs/tvae_300epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 300)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..0f9ba3b84141ed116906722286572ffbc1e485f3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0b6f6f428efd9697992dd2e66e6698d454a85a1182b5565c8df88c5a4cd8cb2 +size 286 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/gen_20260519_094851.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/gen_20260519_094851.log new file mode 100644 index 0000000000000000000000000000000000000000..2bd93654180afb65cafb83b556aea92402bde005 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/gen_20260519_094851.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec20263a047bab029d889cda36ed68f800b184903dd0538a013851a26c2a63b8 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/models_300epochs/train_20260519_094227.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/models_300epochs/train_20260519_094227.log new file mode 100644 index 0000000000000000000000000000000000000000..2f1678de464ecaa4338ac91cb9397e71e5af19f7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/models_300epochs/train_20260519_094227.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:213819013b9391dab42b7fca16f5b1f8946041456d9dd55a8f5eaad31bd2ed36 +size 682 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/models_300epochs/tvae_300epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/models_300epochs/tvae_300epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..c17b56bd4c178a263df672687835284d3ed9581b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/models_300epochs/tvae_300epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:735101b9cd4574ea3dc6e53562e6456e628f8eabb9a9ef4fb3bb6e679d6e7e51 +size 1142066 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4d5de4a41eb8cdc5aaaa2d3f8060dba113b651ab --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6282c42a9f0019242081180b4771cbc5156efcfa4fdb13ef4083310cee13ae3 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..327932cb2b0063b4405fe8a1479e26df26e47025 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e20f606ee58b54d997715e0cebd48a3486c70fce1356a96e5729fa1f28848667 +size 2455 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..707e630a0ed53a2257252780382e05f57f920534 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efec19d215c78336b9500b198ecca784a4c1ad8f6e0e0060e9e214ebe3879ad5 +size 903 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..f9c457f6c427481e6e7cfa8245e508406ebd4f27 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd9ad7fedfd14369be2e45e7583de23947eaef4f6a0b32b168cfe91f0d797e6c +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e47b77333f82e403b6b735650721552b8ac37e89 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b5bea4802583627a741ee48addb8fe4a1d8c35224753c3875d414a8ae536257 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/tvae-n11-15215-20260519_094851.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/tvae-n11-15215-20260519_094851.csv new file mode 100644 index 0000000000000000000000000000000000000000..0aa56d7219771e05b46c39abe9d3ee6d32181968 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/tvae-n11-15215-20260519_094851.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8b7172940048167294e039fa2d4ff2434e3eb62342f93d0d5d08d715892f3e9 +size 2885041 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094227/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..e44babf494850ac8a419f17b477a413be8e1ebed --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094947/models_50epochs/tvae_50epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094947/tvae-n11-15215-20260519_095825.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094947/tvae-n11-15215-20260519_095825.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..2e563a4dd0fb94d3b4261c235d3a61015cbf8905 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094947/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094947/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_094947/models_50epochs/tvae_50epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 50)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..1cbcdb5ed46453341efd451ad9e5df1c01c7b51d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35413420ddc74e342b437b310b25af3b8833b780845f4309570ea306a5420ba4 +size 276 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/gen_20260519_095825.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/gen_20260519_095825.log new file mode 100644 index 0000000000000000000000000000000000000000..cfc16c11dd7775da24c946913aab035e03abe0c0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/gen_20260519_095825.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc9bc09ea699f1e3ef1bec9f230bfe109b214e3ee9f7553f9f4d65f2b9f31fa9 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/models_50epochs/train_20260519_094948.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/models_50epochs/train_20260519_094948.log new file mode 100644 index 0000000000000000000000000000000000000000..4eb5c2c86fa8fa257b70b7c0a873622a6d044ee1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/models_50epochs/train_20260519_094948.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3638c4e06b193b16c8aaec0849a0ff33caaa61101466395e113c01ff96ee73f0 +size 671 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/models_50epochs/tvae_50epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/models_50epochs/tvae_50epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..ab4e6f0dc84a81069cbc6c32d62de6edd2c47ece --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/models_50epochs/tvae_50epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cb30b653a7cc2bd87a95758d9a1fc7e40b63042385e95ec4eb32b27548c87cd +size 1038755 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0fae9cba26b1e4c102d0c5eaeacf0392cd178e0d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8726813cec2f20ccc0dff954138b25840e6eaac9f7e55f11c767f7054effa26 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..1d3c60eb15f1552b06c8ac79a9c00b11667b50a5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70cbc98d034c1aa97ad9d15be19b8aa41e4993900147fcbb1ce23adf88580d32 +size 2443 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..681353779aab506d50901a582f13229147ac84d4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5016ef9e6a0c238b0b931d3100aa47331c8244a9ab8b45e5de5873747248f7cd +size 901 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..bf5983a51ac17b312a99609c75dc1ca7f0939799 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a62826a8b76c499e8dfd1af6f3418812f0059f555a5766d6bbe635614b9be8a9 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..341c233ce6386fc6ae96bc29a54984da46927161 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:950aae71629d8f98b6eccbc0e1ee6e4cc742a7c367d679b9ac1d1dc95eab2d38 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/tvae-n11-15215-20260519_095825.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/tvae-n11-15215-20260519_095825.csv new file mode 100644 index 0000000000000000000000000000000000000000..96d6df3465c98bb7097ec94c5e0d03028aab4d43 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/tvae-n11-15215-20260519_095825.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a403334d85dfcce7c72b9e4604feda63b3b55649a96d5b85c7a61af58d568fe +size 2884392 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_094947/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..0c93db09695dcaadeaaf8e83c3e5d508b894b6fa --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_095927/models_300epochs/tvae_300epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_095927/tvae-n11-15215-20260519_105018.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_095927/tvae-n11-15215-20260519_105018.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..25a52427758f7ffed7c221345e2d7d7b81dd33bd --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_095927/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_095927/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_095927/models_300epochs/tvae_300epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 300)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..75ef89d674c887378e6b877d01c02205d439c398 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f752892d472edb7f41ab9629987e29de921a9b1165d2de766ca2a864063e8e27 +size 284 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/gen_20260519_105018.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/gen_20260519_105018.log new file mode 100644 index 0000000000000000000000000000000000000000..764cd892de6f2347ff45e61b0010a82e0a67df86 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/gen_20260519_105018.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:870f7b6f2fa1b497fa71a86898aba396519fc0004df4ba51e745e7059fc55d60 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/models_300epochs/train_20260519_095927.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/models_300epochs/train_20260519_095927.log new file mode 100644 index 0000000000000000000000000000000000000000..1c7d949f59d2b154e4ece9e1526c29bcf6da8102 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/models_300epochs/train_20260519_095927.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd3161352f2f92d9f9a86f59b53be50ea94f21c4387cd3f0c3f6be21325050f3 +size 681 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/models_300epochs/tvae_300epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/models_300epochs/tvae_300epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..251eff5bf1735304729e058cf9403c8ed0469e3d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/models_300epochs/tvae_300epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2ead138ac86abf36cea49922db014c120d8d0b2fb4e0f89cb5bc7f15c19d2bb +size 4778674 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6e6704db59847d3eb51e3cd4f2b0d1c84527010a --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a423d7c81a7468fb2eb08d8c5e3fcbac372ac370c82364ac11e20e312bac6b2 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..bc484193addec7775dc6daabe598305219f37040 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26b63328ec58c199ad0d07ce0164b05b4d08fb511db595086079e75336dd95eb +size 2453 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..fc42ce96d59f2796613b9fb9b2e0c2e91ea5dba2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21e6cf9c13717dc5f7418982d7efa87ac95b9e666ec8a3365b7e03c9fb754442 +size 904 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..7e44159d5a7d0f378dc663cbcd46524e45135375 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44600fbf2fc61dfb70e44b8593f170f9442dc29198c9319382d25b20ff35786b +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..a5f3af3323f7482339b484aee927419c7d15f5c6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e54ab0291a28479d40b668b215b794c11156e2063118dcfd047ed733195906a +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/tvae-n11-15215-20260519_105018.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/tvae-n11-15215-20260519_105018.csv new file mode 100644 index 0000000000000000000000000000000000000000..e570439b87fad021862efaf1e5c9687d60ce672b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/tvae-n11-15215-20260519_105018.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d2a547a7e481e240bd5a7ed2e9f571860dc19c21d8e5236ce426627adde5709 +size 2884879 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_095927/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..34633d3ee0306c7ae962a7531c2187b7e24d6987 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_105124/models_350epochs/tvae_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_105124/tvae-n11-15215-20260519_105906.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_105124/tvae-n11-15215-20260519_105906.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ab377f7607a4b7fae5fdbd09c0b6186e80e0aa69 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_105124/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_105124/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_105124/models_350epochs/tvae_350epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 350)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..dce69bb54a9ff72f41ee6715cb6ee210c0b701bd --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c301fa052c82be6db24694c9dddec77ec8855824aa4df8e5f8c39ee03cf814b +size 266 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/gen_20260519_105906.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/gen_20260519_105906.log new file mode 100644 index 0000000000000000000000000000000000000000..98f1860e4d0525893c730f365fd150bc2e1e3790 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/gen_20260519_105906.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa13b16fe687547f43bb370706a1a571a582203fb57e691b1adb5f79ce8817d6 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/models_350epochs/train_20260519_105124.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/models_350epochs/train_20260519_105124.log new file mode 100644 index 0000000000000000000000000000000000000000..f601dfdfaf33a2c7f938f8f7ad878d19da136e2f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/models_350epochs/train_20260519_105124.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af3de18b43c4d721fe33b5d7f4cf08f299bcbfd901f62e7b15a027f96272e68d +size 670 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/models_350epochs/tvae_350epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/models_350epochs/tvae_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..8fee57d0842abdbef51f1f217b31c29160594b16 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/models_350epochs/tvae_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2db5d3f0edaee578f29cc7d7c3a2f20f495be4dbc92324bacb196713279e6598 +size 1036332 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..92a7c2afdfce97a56c1de8fd481353413f70255f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a29cda77cecdc68af1d00458ae5ba928b6fe81acc0ba333aaceaee33682a962 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..e59afb73c74d7f93a8c3ce50ec6efb3f6f4c9c07 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a18da038056aab3a05624068d47541daf6ea6802ba5f4753d920f3f17cd3d78b +size 2445 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..e243cd299635a992b61ab9c155431f002e297730 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e685889e627196fef581cd2d323c97eef528e356d7d8b42485becd6865b8d9d +size 902 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..98db036036e71b519388d90244cd36b4e99c8ed6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b60b32497e291d05fdb3a4e13d34d57683f1aad2447d644e039b920e0a15dfd +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..cb8c2cb31422f7e75d6e73da1b57f63384394919 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c173fb2499fc9e27913d74fe8ca3976f16200c77146bb1de2cc61c74927b596e +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/tvae-n11-15215-20260519_105906.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/tvae-n11-15215-20260519_105906.csv new file mode 100644 index 0000000000000000000000000000000000000000..01190a7788bd3740fc2c4581e1312c661a24d5d5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/tvae-n11-15215-20260519_105906.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2543017ff3eaa43b10623d3b2df8059b17e458d9f75508e504086a0b89b3c86c +size 2884846 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_105124/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..8e5aa9cf7e6b33c4ead14f248ac37c5e6e489e0e --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110006/models_50epochs/tvae_50epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110006/tvae-n11-15215-20260519_110125.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110006/tvae-n11-15215-20260519_110125.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..0a3905b9b31e0a497801a5cdcab5bd858588bd98 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110006/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110006/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110006/models_50epochs/tvae_50epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 50)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..3478e589a61d2bd9459f5e4c47ab01e204886657 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a1ced8375f34d20d5daea4d25e535f3ce25c068182b51c1feb769c9f4d2b747 +size 248 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/gen_20260519_110125.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/gen_20260519_110125.log new file mode 100644 index 0000000000000000000000000000000000000000..33369898a9640a9243e4a65aba1d74e517f0c0f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/gen_20260519_110125.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c99f7c460ac043c8a0d8cdc210f998b40336e2b033e182b63f2bc7fd48508eb +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/models_50epochs/train_20260519_110007.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/models_50epochs/train_20260519_110007.log new file mode 100644 index 0000000000000000000000000000000000000000..906b43ca028bf328faab8e11add52327f9aab2ac --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/models_50epochs/train_20260519_110007.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4aaecc9103f07694132e16e26e268c614706c4641228200bb5badfd2563db62 +size 658 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/models_50epochs/tvae_50epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/models_50epochs/tvae_50epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..1ac85acc335e18210b370d600735962089b14589 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/models_50epochs/tvae_50epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d10f5df5728dd78cf55e32dda4074bea14d9edbacd125b51cef5d65c3e64ee58 +size 720603 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9eca9a83625af2bc9da8774add253a8da1c5b0a9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b3c4109b0f978a64382db59c49d73a3fa6ed09c3e055e754e6a39d3e7e5472f +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..d154529ae8a55d0a41b41d0d1bbfa038ba2cd574 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:058a99cd5482c5e21dcdd535a4010de37961fbd718d8cb16168b138ba8684b56 +size 2434 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..a464324a28b45ac200688ebdee71186c77327659 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6c6e06ca903a72a02a9fa48c3c6e2a2a532cb34bf9ba06d1ddff3e4f922bd0b +size 900 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..0971702a43be2e9fe397db735fc12854ecdb9a7e --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63d8dc508771a9eecf9ae52394472798e1887a04ce4bc045dbd0cdd69efdac45 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..7224aed6602881a5f8d16e928afcd4910f70e86b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a76532f8706541b0c3818134b877f494e2cc2652d45c2236919931b5559196f4 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/tvae-n11-15215-20260519_110125.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/tvae-n11-15215-20260519_110125.csv new file mode 100644 index 0000000000000000000000000000000000000000..ad4dcdd350f0832fc458eb404d8de29208aba037 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/tvae-n11-15215-20260519_110125.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cea0b5d992ef448cf67715ce0765126ea0c0ee34159ce326e9322cb88713149 +size 2885008 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110006/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..4765fa81b6183f38aed9122afac4a268fe10f98c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110233/models_250epochs/tvae_250epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110233/tvae-n11-15215-20260519_112342.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110233/tvae-n11-15215-20260519_112342.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..3dd4b2a79623d29b406e927fbc3bf2eb50428b85 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110233/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110233/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_110233/models_250epochs/tvae_250epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 250)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..c219549b6aa24a782a695fa50c8bcec43d877062 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1f31dd9dc89076a08b246d6a706337c87e49a28ddf5b96c3acbfa09e456248d +size 284 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/gen_20260519_112342.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/gen_20260519_112342.log new file mode 100644 index 0000000000000000000000000000000000000000..d39d2c28924dc7beaefe342973cff19107b12635 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/gen_20260519_112342.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23c789839279ceeadf980fe1928a2b6b5f3821f71c8e5577d68cecffb8e1d56e +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/models_250epochs/train_20260519_110233.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/models_250epochs/train_20260519_110233.log new file mode 100644 index 0000000000000000000000000000000000000000..bf824788dc87e548aa67ed3600e90b49e553c519 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/models_250epochs/train_20260519_110233.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76dcec4d421967078ca122b3e062f718da6d3ab89f1d6f6f12c7ed5dfc664359 +size 681 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/models_250epochs/tvae_250epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/models_250epochs/tvae_250epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..ab918fbaa63e68ad27b9edc9253c5e84f5de723c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/models_250epochs/tvae_250epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f300a7f8378c1cfc7f2b79357dc3d165cfb375845468378543b1a0c1aaa37d0b +size 4820530 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..062213a29a5c79551a4868ba9e1d20e1a1fd98de --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8efceadac740e9d30e6b5cb3cb4cf6b590af260d80784ddba4c2b3d5e8f37c5 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..bfcbec74ddc8a67465ee03f49963183ebfebf4cf --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8adf33b1b44d4420dcebc3d673646437c76b943994db4a76f0de9160b9c3ca0 +size 2453 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..69624970718afcf334db0c85305322a93a0dd3e0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bb17cde17bfcdd95cb0c56bfa95780de1d937eed7e670cd1b14e6f9ebf0fde4 +size 904 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..0ff7a55d2fc98f4db4df0ca7d0b6af78ab586ac8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af7ec154b8968b92be41a085e8c051359e8367315c324eedc144493a2e84be48 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c7a696a9673b3f9ee8bed26b4e16983f1afc87e1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65ca6386f3371ea2e6a1bfc02be6fc29b471879cd2a2ed3ad97a6180956ca2af +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/tvae-n11-15215-20260519_112342.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/tvae-n11-15215-20260519_112342.csv new file mode 100644 index 0000000000000000000000000000000000000000..25f693a2e1961d8a89af2c6a497d925da74ca226 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/tvae-n11-15215-20260519_112342.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:354c425e61f3694f2238d38f9ca53d3ef77309286dd059af1a6b547ad7e84c84 +size 2884975 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_110233/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..7e314f1d86ecf59b9d7db5d82f7c9d81d0f948b5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_112455/models_300epochs/tvae_300epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_112455/tvae-n11-15215-20260519_113120.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_112455/tvae-n11-15215-20260519_113120.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..e86b55cd53f578d2d04ceab4995e5515161583ce --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_112455/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_112455/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_112455/models_300epochs/tvae_300epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 300)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..a45952ef2ede26a8856d8f9dd889bbac431989a0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1d9e511f4d07936f69686ee911f49b950697520bfee0028b9dff2212e333110 +size 249 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/gen_20260519_113120.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/gen_20260519_113120.log new file mode 100644 index 0000000000000000000000000000000000000000..188b1d67b3e32b460e8380df41361d811ff143f7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/gen_20260519_113120.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6382e3a57e69f9e7e84d30c856a5f92647a1b2d445c6c02bcc98144cca08c15 +size 408 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/models_300epochs/train_20260519_112455.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/models_300epochs/train_20260519_112455.log new file mode 100644 index 0000000000000000000000000000000000000000..e15723aded94206be3fec3a3b196e39c61283db0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/models_300epochs/train_20260519_112455.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7ea49a41f8cf32e93aaa432bd0187a0116661ebb782d745cf3654e5cb4e773b +size 663 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/models_300epochs/tvae_300epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/models_300epochs/tvae_300epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..29e73d99313453ce7b416bbb8f5e948314d4b216 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/models_300epochs/tvae_300epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1f6e66796f34e885ebd360e9c639c6035fdfbbc4c0ac69c9053d8b7c4fa1857 +size 884646 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..84dabfa7363b818c83e9470f24b093cf2b88f8f0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:204648765e7ab7e09b71f7e0ad55b9e8672cb524deea6c70d49a8fb0de094587 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..171659df9ad7a3f8329cb8c19241c92e88db3a3b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efe7eaef4d175f591373527eec487c64090bb27d626368614f6fdbf07dde2ab2 +size 2438 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..acbb9f5e43690a26bd6efc2d03978db1ade763a1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a120beab3bd206595398a2b0258cdfceff6683470795b5f81db4e09b99102d1 +size 903 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..b08cf35439e5f556e2a36837bf037a3aa54f2a79 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55808f3d0cc6ee48bcb4a1dbb3ee1a3cceb0cdd486281bde1baffa5971bc50c7 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..bf08fc4d8f560e98f38572e4e5be8331dc6d1255 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b317dbd4dd4b2aaeb29e2563a21fe07529f80ede5ded4c54264e090579d129a +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/tvae-n11-15215-20260519_113120.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/tvae-n11-15215-20260519_113120.csv new file mode 100644 index 0000000000000000000000000000000000000000..70b8e1d339bcf0c86deaff700cfbb4c6bd97dbc7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/tvae-n11-15215-20260519_113120.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fa232781c99064151733924d82f4ea78e5f878075bd6e7aec348b92782e2323 +size 2884722 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_112455/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..8ee0e341c74f804b29f4352a4bd31027cf910dfd --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_113237/models_75epochs/tvae_75epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_113237/tvae-n11-15215-20260519_113904.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_113237/tvae-n11-15215-20260519_113904.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..5ed3cd8555e869dfb2eab552526861c844c358ab --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_113237/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_113237/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260519_113237/models_75epochs/tvae_75epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 75)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..215074b792060209518e79f1f3565b45689d6c97 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:549548c1b7f56eb6413632724cb7f037c290c2dbccf0f6fe488c315ec5f6d746 +size 277 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/gen_20260519_113904.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/gen_20260519_113904.log new file mode 100644 index 0000000000000000000000000000000000000000..e004950749ebb1ba84e10af7289284704f1a1a47 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/gen_20260519_113904.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7549bb39685ab426a9a54b8251894eab97b1fd62bfb530525b0a6feb2b0faed1 +size 408 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/models_75epochs/train_20260519_113237.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/models_75epochs/train_20260519_113237.log new file mode 100644 index 0000000000000000000000000000000000000000..0d583eebbb07f868f50b3c45967e13574cc2d022 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/models_75epochs/train_20260519_113237.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19a25ded93ca436fd07cf2e4824bb590122792c4b95d0196eb043a7089b67893 +size 672 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/models_75epochs/tvae_75epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/models_75epochs/tvae_75epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..a9874a31f7b045fe27fd56ecc127412be88a60a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/models_75epochs/tvae_75epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:496ab71bcf2d276dfebfac9e2d478c0a7dc43b03f27b5d57dcb94ec03691777d +size 874403 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..bd10470f9a0755b4ae32ba481cbe9a7b384b0fa0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1873d432f78e38f70770ccce3a78d30f1ca22f91a071ced301a628caf5d1a59c +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..501a11fbff1c42d107072f6be72bb3afb186cbf6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28434d09127e7ff16de6455d8dadfa37212d438e924b88c1a69def938edc175c +size 2444 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..bb0eb11ba4050cda3f8088f8906224771b936550 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a297784077b22b386c47ea6865be29ea01ec0942561453090216a61e1ab5fdad +size 901 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..78c08771aaadab3f6782f3f8b456e2b9efe6278c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04257af5848eeee032c3bd54ed2f3f9fb5da9b2063c17c7a4b9fdbaff9add1e2 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b321de9b7f68f1292451890db0facc2e7fc14e4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c3b21e8f26d8d7f08e66c29dbbefd26ac90766aa046d3bf5cf8eed79c9722bd +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/tvae-n11-15215-20260519_113904.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/tvae-n11-15215-20260519_113904.csv new file mode 100644 index 0000000000000000000000000000000000000000..5573f2bb0afe35dc5ccafb73f53a3cb4d5f5cddd --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/tvae-n11-15215-20260519_113904.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9830cf2620e583ed4a98af259e149812f62437ce4cf99be5098d64c2e9f24dbc +size 2885060 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260519_113237/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..da8206ccbd40d1556b2a85b4a6c42a846e4087b2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163535/models_350epochs/tvae_350epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163535/tvae-n11-15215-20260520_172949.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163535/tvae-n11-15215-20260520_172949.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..7c86e869953b0344cf04a3d36fbee9ebc53a8961 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163535/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163535/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163535/models_350epochs/tvae_350epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 350)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..2f644cdf2925559d22d386fca073b1c83faa1136 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11575612ecf1d2dd44614010ad59b58ea71609ee52dc965f7a890d9860c1185a +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/gen_20260520_172949.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/gen_20260520_172949.log new file mode 100644 index 0000000000000000000000000000000000000000..c40d72c08f3bc2c590250ac8a9b43a2ef7bd3949 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/gen_20260520_172949.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a930f87eb5bb611b72b3b32b981bf36b8df71bd544d680b056ff09a7e1594a5f +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/models_350epochs/train_20260520_163535.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/models_350epochs/train_20260520_163535.log new file mode 100644 index 0000000000000000000000000000000000000000..b2587a6f5f5585b8b3a332c91efe577d529a530c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/models_350epochs/train_20260520_163535.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2d467dd6db69bd03cfe3d73c29908d61f69e9effc910f588de6f585e73c2901 +size 667 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/models_350epochs/tvae_350epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/models_350epochs/tvae_350epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..c38467901484458a101b90435227681dc73fda57 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/models_350epochs/tvae_350epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af7780f0644d96086023dd856a8da10438a58300246cf5b41748c5e86b39e3e1 +size 4889324 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ed616050bde2b7f223b31889ef40a633485994d9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e749abc0b76464f34ada31216fb92eea8293b748a1c7d99bbc6ee2cd95aecc8 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..ae5206e671cbcabb9281b95c9cae5d230663b483 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0003dc4ceb796341bf3e84f275f66ab8ee5044236d7e0d072830ab63842ed282 +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..93f5df9bab34482eb2f6b69025f41ff018ea35da --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57bc533268eca5cb70856cfac2fe29b1deb89ad6806f1c86f90608a1ba77659f +size 904 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..60221f6c2ba36dfcd79d00932af6bb236a8f1450 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:190118d16f4970aa761884345ccd49ab8ab47aba0c742d3a9ee30ed1a1c25bb0 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0ab46f25c7a0ab64ed66435bcbba6e70a6a90388 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe905dbeed1a634c04ee00357007237a0f2f4ad13a2932cc6f228a733a36e9f4 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/tvae-n11-15215-20260520_172949.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/tvae-n11-15215-20260520_172949.csv new file mode 100644 index 0000000000000000000000000000000000000000..df504c3ebc9e3ab45f077dcd43bed7b4bff71ef5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/tvae-n11-15215-20260520_172949.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a7baf0cd9d2b13dbad3c19d56a7cf48acb53c96fb4eea872d83b21bc3b946d8 +size 2884581 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163535/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..f320a4354479addc953ffae482e9f7f6979d8cc6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163616/models_75epochs/tvae_75epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163616/tvae-n11-15215-20260520_163955.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163616/tvae-n11-15215-20260520_163955.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..078abdad35127aecbfe681e8469c8193efed0d32 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163616/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163616/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_163616/models_75epochs/tvae_75epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 75)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..75fec22eb10528fb6be1e2224e04518783cfb0d1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f97192e42919e418322daf283381eab5fdc03099d6f92c79a1996b9df2919d16 +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/gen_20260520_163955.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/gen_20260520_163955.log new file mode 100644 index 0000000000000000000000000000000000000000..adb335a5f837fcd65dccbd1e2e635bd4f886712a --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/gen_20260520_163955.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb9fa97afd54b30efffb6a4eb86a4f52716164bcb315333c32962b357da66a03 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/models_75epochs/train_20260520_163616.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/models_75epochs/train_20260520_163616.log new file mode 100644 index 0000000000000000000000000000000000000000..f17bf237bf3985bed345bf3c73577a218a1c921c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/models_75epochs/train_20260520_163616.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70c718b73c7a2422344b2779ded177ce68501c2c83c42911066f0d5da18457e9 +size 664 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/models_75epochs/tvae_75epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/models_75epochs/tvae_75epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..f080fc821e7f1906f6cb9fd2d79d7cbdca4a5a63 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/models_75epochs/tvae_75epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16f7cd67d5868faf8351a11c3779bce484e5b3f44d33d6c6987bc239709d0ef3 +size 602335 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..87882824301c65f7202bec92b55ef62235bea990 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e2714146f8d856ed92409504690353325f561e5278477e0d4b98489bfe743d2 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..fd52c9eb3a85192c689f2f6c9a90998a8573f3a8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5a9b8066416c594b76516fd890d7d33a65531c4dcfccb3a7a91df0f39eea76d +size 2438 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..764829b26f5c5811a1be267c47278bfb87acc7a7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ee19dd338dcaf4ff75b14d71fad36afeffe4a82041caf70754326bd9d8f2e40 +size 901 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..1f2eae68138bb20bd4c7fc9eaf564eb04e3ef8d1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c74dbb263dcc87e464ec44651a15a19d919b3c3718a58a0fee433bd19647fab +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..124925b36d4e181cd773b0c4e44d587a20e87796 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fc02273398c5af525931663c81d717f0886c15e0c4bdb38246904cde392cdc0 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/tvae-n11-15215-20260520_163955.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/tvae-n11-15215-20260520_163955.csv new file mode 100644 index 0000000000000000000000000000000000000000..778f7651f1c7e9c5fa7edf0a1ce45926dad54b83 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/tvae-n11-15215-20260520_163955.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cd4b3a5c76efb4b5db4b5da5f9d093d2c42e5df567b864393aa5785ee7f636c +size 2884051 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_163616/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..822acb4c10879e5cfc3d6967db9bdacdf013a4a2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_164121/models_400epochs/tvae_400epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_164121/tvae-n11-15215-20260520_165902.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_164121/tvae-n11-15215-20260520_165902.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..4dd9f8ca5df5d4bca66515e426035a35e4ddaec9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_164121/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_164121/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_164121/models_400epochs/tvae_400epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 400)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..36b4fb1644ecf37afb4e2e9b41eb800cca3dd701 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b1c7665b6415825b5f644f04aa4194cd7d5060fecac8b27af5c1c9d91e4936c +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/gen_20260520_165902.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/gen_20260520_165902.log new file mode 100644 index 0000000000000000000000000000000000000000..f045f79bec24a85d5977121c2aa74eee9f5d7219 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/gen_20260520_165902.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5deb6374917436f5bacf7a318f303546e02cec8d20e503e4522f83d4cd9e676 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/models_400epochs/train_20260520_164122.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/models_400epochs/train_20260520_164122.log new file mode 100644 index 0000000000000000000000000000000000000000..3547f33b440eae5d11e5a193e8b802bd7c0005a3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/models_400epochs/train_20260520_164122.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caf98e8bbdfa0552e1d0048278af835cd9e341aece9d9d08a777bbc2ff48d9b6 +size 667 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/models_400epochs/tvae_400epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/models_400epochs/tvae_400epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..a56cb1f14c26886d01608d2be459e58dfc17ead1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/models_400epochs/tvae_400epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ead422ca484bf9365ca73d3fcd0d74d8181361345f5b5bb902667e3a77533567 +size 1676844 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..487e613ee8d16f6bd29ffae7feb1cad7643f8aaf --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:912b8d96ff6540fe3fd7350c4f0b24284fad9ec4007ceb4f71a1b6ffae53f1c2 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..6b1cf73ac119f800fc9067cb34b1e62341b4000f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:643c8bf05b047ac18d0167788be3e99217fccc7751f2a03def1b77c4f97757d3 +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..de7fdb93a5fdd87e58fb16d9dd54e0606a68e549 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8183135fc8846dfd1889763b95a1d55baec179ca2b9b3f1daca076eebe9571ad +size 903 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..a00fed672df6e882777c9ea5cf8639018b07c345 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b838901f3775e08a53f458150f7f071ffae63b57c0979f8eaf34c0a55481fca9 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..78c1a70e4dae397a22527ce828335c6f6483f807 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:faa7788cef3b2cdc44c68abba06a694666ae1b7922e01d3cc876e67e9c63fcac +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/tvae-n11-15215-20260520_165902.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/tvae-n11-15215-20260520_165902.csv new file mode 100644 index 0000000000000000000000000000000000000000..fdb0ca0fb55fef2fa20176ee75401f40c9cabb4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/tvae-n11-15215-20260520_165902.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fef1a311c2444dc05eff48a699b8ef66f7680c574cd61d2c0c49346ac70fb97 +size 2884043 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_164121/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..8de95d4d51f68574186f473beb85f8e36b61b14c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_170023/models_150epochs/tvae_150epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_170023/tvae-n11-15215-20260520_171326.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_170023/tvae-n11-15215-20260520_171326.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..94eb671a80c2da7f7c0bf19c232aa0eff0028d2c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_170023/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_170023/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_170023/models_150epochs/tvae_150epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 150)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..2db447c15bc8c291dde79535bcc9e7088efbe96a --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80a228b65160008fd7c570b690b6c529c879133e9df750be0e70a3875ba0ef6d +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/gen_20260520_171326.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/gen_20260520_171326.log new file mode 100644 index 0000000000000000000000000000000000000000..bce459175b391a4b720641ee74c6c80ed4663b71 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/gen_20260520_171326.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea90412ea31d9d87ceceadd041e032869a570f234e8f3016493bc0fd9f943bb0 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/models_150epochs/train_20260520_170023.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/models_150epochs/train_20260520_170023.log new file mode 100644 index 0000000000000000000000000000000000000000..15c5bc72768a32e3d33063a8afc6c27a254894c3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/models_150epochs/train_20260520_170023.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54721019b9070135970c246d20d1b7faa61bddb2dc1b9b5ac1d12b1df92150ff +size 666 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/models_150epochs/tvae_150epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/models_150epochs/tvae_150epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..763d0cb5f63596544ee2f11a110f530c0cde18d3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/models_150epochs/tvae_150epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eb47d8b08a170b9ccf1a59c4dc9f7cf7b5b5040b4bc9cd1473339627263a5f8 +size 1306732 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4843044b8b6662ebf14773ff4af05f97cce42419 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:584f276132eef126f3e2d8bf837d43de3223a9e84e599a6a84c4d888928c9e30 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..57aa3ca79820caedb4e91e2c6bedcabd2d684e29 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1aa3f2aab0eea0e6e38ae94115cc7c6dc5ebd4f16d6dedc6272bef5ca35ee9d8 +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..e5339dba27fd0c6b30c9321c19dc95994c6000a6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f69e75ca7d763c39fb3085020cd3a118cb076ee8e3bfd2e4b86d3a1151442729 +size 902 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..c261a6a1b48e081c667bd9db9360f2e4d15120f8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e9f4f8bad605c4d9ca816dd7e4761d34633a97e3c0e0237a92ae8938b51bfe4 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b3149ef59eb0bd653f5b8521e71cd0a3a9133c43 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0121c474dfe705af3f044a8414cb713b23ec665f19034aaaf79a272629c8c3f +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/tvae-n11-15215-20260520_171326.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/tvae-n11-15215-20260520_171326.csv new file mode 100644 index 0000000000000000000000000000000000000000..91dd318d9853018d7b0816f287c0d55fcf3ab576 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/tvae-n11-15215-20260520_171326.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5616ed487fad397b616ba9430f9d74d7456903a127054f6fc3e61e72687a0922 +size 2884819 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_170023/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..653629878e6af06c3d1d5bcc006759b202559567 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_171444/models_75epochs/tvae_75epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_171444/tvae-n11-15215-20260520_172702.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_171444/tvae-n11-15215-20260520_172702.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..fff5dbb0aad02468f2ad3e4eeb5f6886105afcf2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_171444/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_171444/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_171444/models_75epochs/tvae_75epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 75)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..4c142d2e9df0d1e61ddc3315a83356850573fa2a --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5c827356b009609895c6442f46f5cd7e6fed7731f97e6a2700da1865172fa43 +size 244 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/gen_20260520_172702.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/gen_20260520_172702.log new file mode 100644 index 0000000000000000000000000000000000000000..954fbbfc564da30e480b9e4c9204d799fb72004b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/gen_20260520_172702.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:461c6ed69315f970b0a823e206c01269ef47c227bc6afed5da708c1ea67e0a7f +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/models_75epochs/train_20260520_171444.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/models_75epochs/train_20260520_171444.log new file mode 100644 index 0000000000000000000000000000000000000000..91f840667a57586613cd61bf2b88481a2168ded1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/models_75epochs/train_20260520_171444.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e81b8aaa88a6a01b225393884242c31d0be49123530e44b4d9389232b8c1f63 +size 656 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/models_75epochs/tvae_75epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/models_75epochs/tvae_75epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..e34a173355de4b952eb8f72fd4c0464aa2f67d53 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/models_75epochs/tvae_75epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4b6f94f578808d6f2499e1253a3ea574a8b637d2ee7f6c9a476c8f558d552e5 +size 1343195 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d5b1e24e36d1afd8c2c8470d8a82d48e341ac432 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd540d83f76252c8b48024cb8ec5b5b1d351c1707300fb8c35665484f42240e4 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..7e38a79bf844b305ff3dc20ba02a4818beb872df --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b50e9b757d7a02134b17271be2563a49432c8196b77794299ca9cfe117ebf1aa +size 2431 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..4ad21ad854480d89900c5ea9f68d467b8b9d6f3d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c4511ee43ffa76749689fd236dfea99a7220f3cc0996d928a7bb54ca6051230 +size 900 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..d5b4beb1dcd6b31c8641dc555855b959ec963847 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0590bb133f86b68b86dda9dc6186fe6a6e5d4b0c75e2005f5dd9641fe3967140 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4feb227ce8ea6d7f61981b5485c10556498a14c3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:698d4e7cc93b4872ffb7617272103ad4f97b2f15c28a64b5bf23baeb240cacb4 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/tvae-n11-15215-20260520_172702.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/tvae-n11-15215-20260520_172702.csv new file mode 100644 index 0000000000000000000000000000000000000000..084a8faf893b26166efc625829bb021959e7c03b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/tvae-n11-15215-20260520_172702.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85addaeb0faa00b4c47253e7db33ffe1743c09b1c2b7b7653819d798cf8b7e4e +size 2884547 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_171444/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..f50645229798ab220645d5290efeae77e7086a79 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_172821/models_125epochs/tvae_125epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_172821/tvae-n11-15215-20260520_173901.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_172821/tvae-n11-15215-20260520_173901.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..c47e68159d54bfce162db6a3613063d2c072ad01 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_172821/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_172821/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_172821/models_125epochs/tvae_125epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 125)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..2f644cdf2925559d22d386fca073b1c83faa1136 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11575612ecf1d2dd44614010ad59b58ea71609ee52dc965f7a890d9860c1185a +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/gen_20260520_173901.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/gen_20260520_173901.log new file mode 100644 index 0000000000000000000000000000000000000000..f8215c1f4b12de41524cf0989eaf6bc159a1cc3f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/gen_20260520_173901.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53a3d365dc759cdbffe244d58342745bacdf7cae198936716456ec48213dd8d6 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/models_125epochs/train_20260520_172822.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/models_125epochs/train_20260520_172822.log new file mode 100644 index 0000000000000000000000000000000000000000..fae5f41c2c572283aba7acc85a06cd6dc0328577 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/models_125epochs/train_20260520_172822.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:718e862ba1cdb2c578b4a900e4dfd25b2ce424a2a9b4d11b7bd64ee2a154cb89 +size 666 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/models_125epochs/tvae_125epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/models_125epochs/tvae_125epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..d2f3ede04322fb81adaf345b26178dda83fd1bb8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/models_125epochs/tvae_125epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f686d095bb2dc7ffa5a84b7c9114262160fc14e6f9a00b9a44e22a1e2cb4add +size 1176044 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..317783838e7264aa21058358b05560cad2afc9ce --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5fd8d58ce7da5d48785e300b0169408bb0e3608e0746a56a286c1864e8a4e1f +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..b3ef92fd900cdbed70d98292921f7e8438f584e6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79148d24f5fc938547a5ac917175c27ae6ccb277e3e431df3619e54f35cc5e9e +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..5f6ae802f9da3e8afe8d2e1f1a8a937ef3cf5398 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:318b6d6660e6bb4a8303744684fec923bd846482f8193565725077cf6c219a01 +size 903 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..51867a46c75d2312ae09abe2619412325bf43c7e --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:174705f3414337c00f0980943deccdbaa2512d40b1d188de4e431ac65699de7e +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..5b52730c4a5a82507533561d24af6abf507051a2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:230492d74d27ea48e3f59fe2de549024f94281008f3e927c3852931712535ffa +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/tvae-n11-15215-20260520_173901.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/tvae-n11-15215-20260520_173901.csv new file mode 100644 index 0000000000000000000000000000000000000000..c4676c36b1b5aa6c803f49b86ab6c9fc9cfefbfb --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/tvae-n11-15215-20260520_173901.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67e17390248082162a9eafdc61a45a4a8b9db3c2af018eb8b435d2c652d7ba1c +size 2884255 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_172821/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..457f45bf67477acb2fa7dc713cb42476ccd1bd85 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173119/models_75epochs/tvae_75epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173119/tvae-n11-15215-20260520_173444.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173119/tvae-n11-15215-20260520_173444.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..9b4e3275f5657313eb44b702c80b8e9ef54ebe30 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173119/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173119/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173119/models_75epochs/tvae_75epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 75)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..75fec22eb10528fb6be1e2224e04518783cfb0d1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f97192e42919e418322daf283381eab5fdc03099d6f92c79a1996b9df2919d16 +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/gen_20260520_173444.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/gen_20260520_173444.log new file mode 100644 index 0000000000000000000000000000000000000000..9d88035c9f72548e7a04f47fb783e80b501a1205 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/gen_20260520_173444.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e42c58375ad4c0fec904e7d0b7def1993f71da34f2d27d4e9f9f9403c99b9f5 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/models_75epochs/train_20260520_173120.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/models_75epochs/train_20260520_173120.log new file mode 100644 index 0000000000000000000000000000000000000000..1a19b59ea2c8f640dbbc086c426a79724e04aa93 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/models_75epochs/train_20260520_173120.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11207e91536cd73fe795ca86aee46fa84b2d09109fd32256557bdb68712027b4 +size 664 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/models_75epochs/tvae_75epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/models_75epochs/tvae_75epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..df00082d88df9b56addaa6d588ff96b3a2f5156f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/models_75epochs/tvae_75epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7428f81ee557bd5abdd0e8d2aad59c6c254d79fbc89473d7ebe2ed9d3ed4d72 +size 602463 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..703b235d49a3ecad2cd8a188c687aa11b49a8c20 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc18bedc766890df492d07033e8ad238912d259ce2cceefbe5289e469735890b +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..e9469236273549aed7f08a53e59a0be377d04c4f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ec6a30f8698ee74a16c29be2c30859dae78005518408481e588206f58a77b3b +size 2438 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..62c876b42e600ac3202e922a0e8e4e04d7171356 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d933f559a71749d76e3f2a4a4bf27ee6ada862d0bc6a267a1008f74c5427709f +size 901 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..b10de43d193a970a12eb75ee1793330fce36e387 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09c99137d66b92fbf5f519f5a79c199242d1e36962c045ad3cd962c35a828859 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d00a424fe846b1480552bdc8e4a71db377dafe98 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bb4a6849b40e1ac37f77fce91c205581c7647363873b4331667cbdb4e8f129c +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/tvae-n11-15215-20260520_173444.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/tvae-n11-15215-20260520_173444.csv new file mode 100644 index 0000000000000000000000000000000000000000..41dfe79b4cb4f06b14d17da1d3df87712c184a83 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/tvae-n11-15215-20260520_173444.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b82bd1a15bbdf84397777b1a5faa7d29d3db31f7453f178050fcbb0f525c531e +size 2884932 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173119/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..8ed00385f95b2f0beedc33635c740034dd1bf290 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173615/models_400epochs/tvae_400epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173615/tvae-n11-15215-20260520_175306.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173615/tvae-n11-15215-20260520_175306.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ab45f30e2647405a9fdc3097b7b2968e1d37bfc9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173615/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173615/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_173615/models_400epochs/tvae_400epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 400)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..7c0527d4c243a4ad6619a173d38eb870528d442c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c779ce77d59f22d3b14c0bc542896668882288ec696a87bb3dc049b62862e80a +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/gen_20260520_175306.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/gen_20260520_175306.log new file mode 100644 index 0000000000000000000000000000000000000000..7125396315af925d14ebf79ce0bb3406a00dea95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/gen_20260520_175306.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29e2b4c4d7c2be2824bf826095baad999a49406d45ed1bef2051299390fece32 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/models_400epochs/train_20260520_173615.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/models_400epochs/train_20260520_173615.log new file mode 100644 index 0000000000000000000000000000000000000000..bb1eedfa4fb6f0429ac03fb29f7609083f538997 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/models_400epochs/train_20260520_173615.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f014b2021bd26ae5b7a52ea3dabafce517a96aefac9ee69423f682ac583f1b56 +size 667 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/models_400epochs/tvae_400epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/models_400epochs/tvae_400epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..914ecac59b69c5d529db1e51c10e6d5eab5a8c83 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/models_400epochs/tvae_400epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:382059988e53b5377b6b562996cac55e4b46d3f71aef8723a70c7b7c3669235f +size 1676268 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c7b993925e8bd2f481f0e30cddb8768d62afa852 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b836c467da5a942e10e63ee6865419eafa7b0f75f66f3f375d6843cb31dfd586 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..7cd2881210a4c62512e371aa743addbed99eee6f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cd59372eed428cdbc3ab0a7956a1fe000e2dff27e2220a1c602674d4e888230 +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..1422c91849166245e1e86c551bb805f152b4a7d9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9561e46f7dfad7519d33addd45855f3abaebc773586bd4d9e3cdf41374119190 +size 904 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..2eb040c9c5f1015ecdd567e9013208a20b2b77d4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:099d9aee2177cbd9a3bb8eaf5a2681f2aae9e500f461d13030d9b413bcbd05ba +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..00a153d4ff95dd563717ecc2288281e12c0c3193 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37eba2e6fa968b6c30cf3be84c8fa0dce69baff607b506644a56c536eba7a415 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/tvae-n11-15215-20260520_175306.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/tvae-n11-15215-20260520_175306.csv new file mode 100644 index 0000000000000000000000000000000000000000..62a05223377ba2a0ceef38442a6b180208172097 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/tvae-n11-15215-20260520_175306.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f885e3bd61ff855bb82c4229e35575a5e0de921c7cd3d33b42013722e8132da +size 2884305 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_173615/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..dfc4807b4f3d0eb591d36eee37f3ea2100d778b1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_174051/models_125epochs/tvae_125epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_174051/tvae-n11-15215-20260520_180159.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_174051/tvae-n11-15215-20260520_180159.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..44d294e552b976549ee2b1d6ab528a2adc1bae60 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_174051/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_174051/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_174051/models_125epochs/tvae_125epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 125)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..36b4fb1644ecf37afb4e2e9b41eb800cca3dd701 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b1c7665b6415825b5f644f04aa4194cd7d5060fecac8b27af5c1c9d91e4936c +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/gen_20260520_180159.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/gen_20260520_180159.log new file mode 100644 index 0000000000000000000000000000000000000000..9a7ae2fa2039110d1622bc708c5421bea5c73ede --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/gen_20260520_180159.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02cd07cf50a5180c19766560ad4c0aa82b8d799603021ba32dbe4638daaa89a6 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/models_125epochs/train_20260520_174052.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/models_125epochs/train_20260520_174052.log new file mode 100644 index 0000000000000000000000000000000000000000..ee8b023a82a1563baf95cc19a3b35a336744f144 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/models_125epochs/train_20260520_174052.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72bf24581304fafb5924e887e3caf0e4e81ffe11191ebcf9a26c488c56b232d1 +size 671 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/models_125epochs/tvae_125epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/models_125epochs/tvae_125epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..5812b8c4e7d088a24b0cb750dfab272d3a5f78fa --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/models_125epochs/tvae_125epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c7061b73a8027c90120450cc8dc34830d866a68c31d4256858b8180a46c5016 +size 2074092 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..f705baee99c2a830b77e6ced9dc9b30a11c5aa22 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d5ec97d73c7ed276e81f68d8c97c6bb869930e444eef4839da222d0e18f41ea +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..ae459c2ed00981dadd6d24bd6b71307bfd1d0e82 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c77ada855ea821533093b94a6c35563b580d248c713755595f3c32e4cff10782 +size 2445 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..21e384c9dfb19b8c00ad5c6d12b554d5c8553652 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d19862cb46499d14acadc4eadf9f0662bfa34313edc9d1fa2a75f9e72fa8679 +size 904 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..4e4342cef3650f31d24bdfce6f596300c711a2d6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5cc8329e4d563480e2bbdfd1dcef452cba01426494d3b7e4ef09ba4051b53e9 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..f6114b08d3eac00579675d42080b95e3a58831d6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:117a4b791cc28f7295fde5e6dac0a93a440a83d9ec94aebc15f49b472262e082 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/tvae-n11-15215-20260520_180159.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/tvae-n11-15215-20260520_180159.csv new file mode 100644 index 0000000000000000000000000000000000000000..9c3b7d06c16ba1ab43a9af5f0c96e8dc4ff770cf --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/tvae-n11-15215-20260520_180159.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10ee87638a6f1cbb90841bd98dd580533cfeeaa9c7cd9cd49548c1c617295f86 +size 2884326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_174051/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..e534eb944d2382992cac6bfcae055ce3e9a70e52 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_175428/models_150epochs/tvae_150epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_175428/tvae-n11-15215-20260520_180650.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_175428/tvae-n11-15215-20260520_180650.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..d7b34d5621182213ff6614218b22635caf6402f6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_175428/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_175428/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_175428/models_150epochs/tvae_150epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 150)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..96c6efc4742ea8db70fae3ba686916404180cad6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc369ac2fcad1d109b51efa3f7775dee5a59d7b86192ae2bc5bacd04822f5824 +size 266 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/gen_20260520_180650.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/gen_20260520_180650.log new file mode 100644 index 0000000000000000000000000000000000000000..2eff8b9f752b4845d4b9b13f81ac50d121dab0ab --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/gen_20260520_180650.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ead9a57e235fedb7f5b4b4979abf1ac5c916a39bef5513094dc753e645fecd40 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/models_150epochs/train_20260520_175428.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/models_150epochs/train_20260520_175428.log new file mode 100644 index 0000000000000000000000000000000000000000..e9bce57c66a6d3990909bb186eb92cadc6df4785 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/models_150epochs/train_20260520_175428.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6e7b547995ec20f308d954809ae5c27cd95a09191bb1a1c444bdca1a892cf5d +size 665 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/models_150epochs/tvae_150epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/models_150epochs/tvae_150epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..2998bbeb184de11946c3b2b3849bb6f3d8b4fffb --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/models_150epochs/tvae_150epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0c2652db630fb8af409b3ce93b883d52533caa6d14df87a9fc3f3ead02ad55f +size 1306988 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2e6f7cf45fe0f58a0d3707e328cf635e8a85349d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4122f73a366002b1cf99dd783c9e3800b4b20b595d1707416e1efa5093bf478 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9e46bf875d4d0a98c7350a29f06ef4671ebaea07 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c301d9a605e3e119478250ab5002845e6a732fdce176b1abc9ab8bb9d8b9a33 +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..b58d8895c028df463880478ed91b7f068b378ba1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a47423cbb84513ff836873544e2c9dc341612194fcbdd55acb43b184b779915 +size 903 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..cd0af8bde3115c6fd1eded427f82bc59e818697f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a26619bc05e9d567ca643bff4dc82f57358d17ea1ed561de1e7c6444546d2d3a +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fdfaaa199e8497a5ca91d5c4c757c6909fd5cddb --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0e6259e88643b78ce0feab5d90c86c44a718dd84edb64538a4a04516d7afe7a +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/tvae-n11-15215-20260520_180650.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/tvae-n11-15215-20260520_180650.csv new file mode 100644 index 0000000000000000000000000000000000000000..03d09661092d44ac219e174860272156c4f00a37 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/tvae-n11-15215-20260520_180650.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56cbbe0ebc72d94ae18e14ba86772a5829ad0e533908afedc0836a953c504098 +size 2885230 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_175428/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..f55bb166022601f4cdda6cb420a252fe56298612 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180322/models_75epochs/tvae_75epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180322/tvae-n11-15215-20260520_180955.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180322/tvae-n11-15215-20260520_180955.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..bf8b6340d7aa107b9bf7a0172b82150040e81060 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180322/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180322/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180322/models_75epochs/tvae_75epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 75)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..2db447c15bc8c291dde79535bcc9e7088efbe96a --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80a228b65160008fd7c570b690b6c529c879133e9df750be0e70a3875ba0ef6d +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/gen_20260520_180955.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/gen_20260520_180955.log new file mode 100644 index 0000000000000000000000000000000000000000..de102ebb66563b7a8d8b8f2d942996c45a8532f5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/gen_20260520_180955.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3b81875de3caefb8d82404fe534ef373f3383edbe88ad7813b6b0b0d39f3dca +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/models_75epochs/train_20260520_180322.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/models_75epochs/train_20260520_180322.log new file mode 100644 index 0000000000000000000000000000000000000000..a9d41807b31924b4d5a7bf9484a2239e0111e3f0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/models_75epochs/train_20260520_180322.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb8766c2ca7ba6ff20cc33a96bd61f54a9dc58c95f62e23c486e725f767650b9 +size 667 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/models_75epochs/tvae_75epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/models_75epochs/tvae_75epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..83ec4885b43573d9731ad652ccf4e4e7e4fc499c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/models_75epochs/tvae_75epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcbb147f5ddd07d25a55999db40fe117f491be50cc025b0e7d5d366b6a35ac7a +size 955871 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b7736a7110c3f07ffdab25df54f492866816f0fd --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f1e7aca45b990bfde8dc984d48df10cb83f1e59b451e5a1ba251f22329f1df8 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..50243e217c7105adc6ff6b12a4cf9e0a3ec56138 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbacc4cf3e5d1895a4401ae523305dd3cbab128029512ca5c1b92efc1f219ffc +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..aebacb7767faace010854a67db32227ad443aa92 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef78356c43b4efab146984d245a95cac3efd99a6625ce0ad98fbacd632ef71d0 +size 901 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..f76bd72e7d247613ad551a9ac90073f214f3d6b0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cf99cc9e6ca2e321a77a816bf40504e6d42c0435f74c75cf1074f007e5a0304 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9686f418b63dd61cd6c7188624de2f2ee69a006a --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99198b4b63bee4b2abc94a9f78920cbb2aa9c2510f25d937e5235cd611e21da1 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/tvae-n11-15215-20260520_180955.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/tvae-n11-15215-20260520_180955.csv new file mode 100644 index 0000000000000000000000000000000000000000..ce5efb5c84bae505e974986798222d8113f48f66 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/tvae-n11-15215-20260520_180955.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66425aedc524dbbc9dc3b1e71a53a5a9bd137a06ece5e4f47baee70c2f597005 +size 2884302 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180322/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..e3fee81b959dee3dcd46ceca524186127ed55493 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180806/models_75epochs/tvae_75epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180806/tvae-n11-15215-20260520_181929.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180806/tvae-n11-15215-20260520_181929.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..b3633dc70346ac3b974e92efe1d95134d3db8f68 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180806/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180806/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_180806/models_75epochs/tvae_75epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 75)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..f7e851f332c917d934016a36c6414a1536ff63a4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f5069d228d0e4e45543dfae17f5df635d2b5cf86b394a15137c68d72c026387 +size 265 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/gen_20260520_181929.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/gen_20260520_181929.log new file mode 100644 index 0000000000000000000000000000000000000000..00ce1406e021e571ff832d31b45eb11282e484de --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/gen_20260520_181929.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9057b100046ee6c9dbb7b8f4380579afb68af1290e07c69384bd3b4769a4c2e +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/models_75epochs/train_20260520_180807.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/models_75epochs/train_20260520_180807.log new file mode 100644 index 0000000000000000000000000000000000000000..9c271b989c7db531e362c7072b69de0268e29d51 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/models_75epochs/train_20260520_180807.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed6bdf03a6abfb1394b89917dcf307023c8a2638ce064409a4fa83726ff401a3 +size 657 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/models_75epochs/tvae_75epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/models_75epochs/tvae_75epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..63c5c2ce206b4681d72640d123e0f420f4f507ff --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/models_75epochs/tvae_75epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50aa0b968f32327b251dcba64e326032ede59d8761baae22cf9676cea95aa047 +size 1343387 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..7599874e9607856d80cc351416a0a046e97fdb5c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f13242c50269e3785e71a4b118fbde198ea6c7f6ed285478c2052960392c1248 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..dcf352f4e17380ebb5a67ad58fce870438162e47 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1231fd90f1c058063a7304156cefdffa75eba8c30bc7da6058beca9b08ec78a +size 2431 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..627dbf41124d0f2ac204d53616e5044b4f9551b2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49529f1b63b92e179d35fb41fda9e7f6c6c2257cbcb5bd8e391d2f8d6ee8c732 +size 901 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..97a14f1bd9e4e56347a29deb15ad02f6ca147588 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c3c75f24d77f61c7b09ed1822c05ae7715074f60b1fcc7259af142c2b0ec56c +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c9e0ee68803f08141ce85b1b3b51a0beeba141db --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e76fa385bf69bbc76a61607b4194361fbec83509fc6cd978ae6a18052e56c45 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/tvae-n11-15215-20260520_181929.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/tvae-n11-15215-20260520_181929.csv new file mode 100644 index 0000000000000000000000000000000000000000..fb6e6a530a3f1de09a2c63a3f825fb8b15edd607 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/tvae-n11-15215-20260520_181929.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0cc7faa228e2fab598a543e9509c415f44c03c9de57dcb807d3d43aeb9224087 +size 2884237 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_180806/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..1e07a15618a0aa77384bae6f7eccbcc7b7b4807a --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182045/models_125epochs/tvae_125epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182045/tvae-n11-15215-20260520_183042.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182045/tvae-n11-15215-20260520_183042.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..4a92728d4a7fcd28930d6c498acb32c7b072e7c0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182045/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182045/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182045/models_125epochs/tvae_125epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 125)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..e0a553031080d6e3aafd05defa5efd78f7594c9a --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:565273f218660439adfce680de0c5dd47ac78c0168501264f90b50c747ad3d9d +size 248 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/gen_20260520_183042.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/gen_20260520_183042.log new file mode 100644 index 0000000000000000000000000000000000000000..7cdd2d28c1217e2b12c95be28d395767e5287eb5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/gen_20260520_183042.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d680643d1e680ce462a70cff620c48c7f40272c6f5f071056d52df6954490ede +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/models_125epochs/train_20260520_182045.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/models_125epochs/train_20260520_182045.log new file mode 100644 index 0000000000000000000000000000000000000000..8f83393c79a72b81dbe1f0e89d2765f22bc52a82 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/models_125epochs/train_20260520_182045.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f7ff01fb6dd81d4860c0408298140feef596950b80c2bb572a37a6bf3d4abdb +size 666 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/models_125epochs/tvae_125epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/models_125epochs/tvae_125epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..862fed4125a127e88ece8d2ada2e706254a0f36e --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/models_125epochs/tvae_125epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1fd35c181c917741b712a4fef0210a444eced5d535a9932eb0578c34b79f797 +size 1175468 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b8b8ef5abdcd74bdce1ad183e5e96dbe46d0baaa --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7960cb6414942d21b9467c6e997b1eac2ee2a8b55ca2b2d3b104c75cad741bb0 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..a5ea4c91cb0183df603cefb470fa8d8d83032d17 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61ae86f74a920ea38f2939a8ba544c264e5c1d01ea21142242c008fdd1c9e71d +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..d48ba105e296cf71425c236ccfe0b16c698f0f48 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80a03d5a8f30d989fb69d2a5daad6bfb62fb38ee1549dc59bfe64870eec9f6fe +size 903 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..ee4500d747b5bafe295ecd542d735b74cd895d2f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3386202c16f15bb110cd43249657017f7ffbaa3c31f13498e54b614f43702df +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..75c9772e1c110fb42c16bf28984bb71e407ea162 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4501dd2b52b4c07c0b120ab6c25178ad0bc2fd0add2143e3f26a180b1413a7 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/tvae-n11-15215-20260520_183042.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/tvae-n11-15215-20260520_183042.csv new file mode 100644 index 0000000000000000000000000000000000000000..f3b7df3273a7b0ef23df6de89cba4a66856ea475 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/tvae-n11-15215-20260520_183042.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b442016e2ff535a9807c5b95f50fa2ec3dbbf17a0ca15dc9efad9cdaad82157c +size 2884561 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182045/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..6ecde71f68d2edea0905a2dfee83b844471afff5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182256/models_250epochs/tvae_250epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182256/tvae-n11-15215-20260520_190534.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182256/tvae-n11-15215-20260520_190534.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..6712d77c75042a2edb8032bb7647816fa961613e --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182256/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182256/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_182256/models_250epochs/tvae_250epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 250)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..7c0527d4c243a4ad6619a173d38eb870528d442c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c779ce77d59f22d3b14c0bc542896668882288ec696a87bb3dc049b62862e80a +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/gen_20260520_190534.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/gen_20260520_190534.log new file mode 100644 index 0000000000000000000000000000000000000000..bae99a9567bc1e20050c135af9136ff20c56d538 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/gen_20260520_190534.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f6d9dc9f0c0649dd10d65f9a12cdf60ee82cc75cfa51b6a37d1bf3379042134 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/models_250epochs/train_20260520_182257.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/models_250epochs/train_20260520_182257.log new file mode 100644 index 0000000000000000000000000000000000000000..c298d46c3fafb35362772fea4dee1dfd130743f6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/models_250epochs/train_20260520_182257.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:771c80fb5dc8df2062bdf0e7718665eea1cea1c29fb9f07df8b189e1944ba46b +size 667 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/models_250epochs/tvae_250epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/models_250epochs/tvae_250epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..d5eef51085850bb0822b91d2be11efbab8f6c882 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/models_250epochs/tvae_250epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90fa2eafae88a0bcfee9262a463e4983eb39b749a0a2a0dd085cd6b754a25d39 +size 3627116 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d0e2bfcfabbeebc8852fc0870d161f54d20b6f6b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc39d857e09e86d7bbdb58634672e559c4cda587d06dd3e64498b93a8edd6326 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..7f87f7a4a84e7dc6406a5ae9b934d0670bc16575 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66460f7608b1312cf303c644a3ee26f952034a82e0ece719310fa95fd777f6ed +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..0ab88c586d87c4394bdcc8d2c8bd339d2b4b72cf --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d777b45d78b367d972c177e3c5b98859152542c7e28b2502130ce6edd808699 +size 904 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..a8a0caca9b596ec8a87ebb98a9e98eb71a0effde --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73463fc1ae64a497854e4f63f3b6ab2c3d43eb738fa534f052a5879ea0a057ec +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..650820f7c895ac0dbfde49b44b5b41f0057ad11b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76404a9b386b36bc5e9b991ca6395f4e643ddfe23f7b07f239e6c39a0d182348 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/tvae-n11-15215-20260520_190534.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/tvae-n11-15215-20260520_190534.csv new file mode 100644 index 0000000000000000000000000000000000000000..6ef21e4e62fea4007a2164c9db24c1f8e173e10e --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/tvae-n11-15215-20260520_190534.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7478898723e1efc2a2605ec50c4cc922ac3f90c7e4c31db26cea0196c03c7e6c +size 2884568 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_182256/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..56abe52db5bba6cd126e5036e7d6c3f39279cba7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_183158/models_125epochs/tvae_125epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_183158/tvae-n11-15215-20260520_185157.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_183158/tvae-n11-15215-20260520_185157.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..d55415071822f19808a304ef0cafbb22792b69f6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_183158/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_183158/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_183158/models_125epochs/tvae_125epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 125)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..96c6efc4742ea8db70fae3ba686916404180cad6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc369ac2fcad1d109b51efa3f7775dee5a59d7b86192ae2bc5bacd04822f5824 +size 266 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/gen_20260520_185157.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/gen_20260520_185157.log new file mode 100644 index 0000000000000000000000000000000000000000..fe4d2db01ec1279a86e3d4722e8c92dbd24f656b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/gen_20260520_185157.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e5108c7898d8e62669e9b1c9bb4a7a0e1d8284f21626eef7b2b1872674ade13 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/models_125epochs/train_20260520_183158.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/models_125epochs/train_20260520_183158.log new file mode 100644 index 0000000000000000000000000000000000000000..93fb4884a580f98f754a792f597d825db9b506c5 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/models_125epochs/train_20260520_183158.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5be01ee90da0b68ff02da260ebdf9534d2db496f4b90023ccef3ff63a3936afc +size 670 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/models_125epochs/tvae_125epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/models_125epochs/tvae_125epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..0e58736e0d3f2edd9c12c4a16e849e61929a1130 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/models_125epochs/tvae_125epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ff838dd8d34cf592603014c66b434ee63186a89ee2251fe41f009d75bc7ad34 +size 2071660 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..3a1973c02deae4477fda07e7dbd0b0ddcad5110c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d92c2ddbc5997f505884cb655115eb43c3003ae79353eb945e89fcaf87127a1 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..668b6de4c3b1e2f6c972d6bb179cb6a70c5319a2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:968fe3d413f686e11b56f34935b9e165815fcfdc6684faf982e0b869e722070e +size 2445 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..71ee00604688f7a14c58c15f5e7e481d3448e167 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03dbd50720019162e28c42d058467944a7aead3504c0bd96483fa52db88d76cd +size 904 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..b5d8fc61ccc7fc50c4ca221ab6784874f460adc2 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d40a9b6d67ef7419e3bbbf27014408eae143011a2502b3767b587c0601ea6016 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4ccde608e2494615e76ae0a32f9cc57bfcada69f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62fda8bbf7959926ece0ad5137285dcfeeae49fc99b56b2702eb753bee151b5b +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/tvae-n11-15215-20260520_185157.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/tvae-n11-15215-20260520_185157.csv new file mode 100644 index 0000000000000000000000000000000000000000..72102e99ecda75395918c7cf80b88553a13e716b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/tvae-n11-15215-20260520_185157.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8de06b095bf8f349a8496583f68eeab9baff3d713360c185b1f565f42dc15636 +size 2884684 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_183158/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..8f838e764fdd848bcc1317ea40565bf76b7df6c7 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_185317/models_75epochs/tvae_75epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_185317/tvae-n11-15215-20260520_185929.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_185317/tvae-n11-15215-20260520_185929.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..ae4c61321a49714a47a486ac303d91cfdfcf112f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_185317/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_185317/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_185317/models_75epochs/tvae_75epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 75)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..f7e851f332c917d934016a36c6414a1536ff63a4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f5069d228d0e4e45543dfae17f5df635d2b5cf86b394a15137c68d72c026387 +size 265 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/gen_20260520_185929.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/gen_20260520_185929.log new file mode 100644 index 0000000000000000000000000000000000000000..23dff4d46e0135c0695193ac2ed11ce0ac37e5eb --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/gen_20260520_185929.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1299bf92ddd7e832dfe33d15dca6ce8314886cbc87c932f305c73a26fbf97e2 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/models_75epochs/train_20260520_185317.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/models_75epochs/train_20260520_185317.log new file mode 100644 index 0000000000000000000000000000000000000000..9164fad14b3859ae77945b120053128627f61c2b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/models_75epochs/train_20260520_185317.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e8af49fe79abd01f24ab7d4dc415fab286d429f38ba4b9fcc165add1361c7fa +size 667 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/models_75epochs/tvae_75epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/models_75epochs/tvae_75epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..52de6ded6a47a44bacc09d0013c430677ac57427 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/models_75epochs/tvae_75epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88210f3482e160cebaa8b77a32f19337fef47403d7ca8e7377e119f9a9c66ece +size 957023 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..881527919c54a95f542b9c6ecadccb62abcfe97d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ffbcaa164bd17f428a04c7a5072d7179a35e8a165da8c255cdb38aa8a4bac5a +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..3eeafe2b3ee40defe309ebf33c9b2a99dd189ef3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:145de7b8e7ea4f4e36f823b017d4bbe9e769b9595281932cba82e95a1bd9ba5a +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..5c35cf68155649e8760f98ebfcdc7472d0a03853 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ab9ee9013aaa336370111fb81ad199745a2636e164a0a72d4703d09bd8decbe +size 901 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..93269baffd980b4b9bfebe429c8119b11e11483d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08efc4064d7875d7b3f594e90573c31eeffb68e60a16dd475a5833604aad9966 +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..54aba6a46a648cfccde61529b5acae4ca801e95c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90cd5eb587f22b6007f48941a9bce41e1fba1c36451be5ce51eff0b8b6631b84 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/tvae-n11-15215-20260520_185929.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/tvae-n11-15215-20260520_185929.csv new file mode 100644 index 0000000000000000000000000000000000000000..a57fffb03190024b805bfd5cc02834d6745fe204 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/tvae-n11-15215-20260520_185929.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2977ff489cbd1524d72d9253bd65a1ac08c67a6194cdd42b7fb3acdeb66ffac +size 2885376 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_185317/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..206312753f833b6f6a8f0e27b54ba5af061eb0ac --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_190045/models_400epochs/tvae_400epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_190045/tvae-n11-15215-20260520_200013.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_190045/tvae-n11-15215-20260520_200013.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..0e311e02963883a512581c5842b3a9fbd6267cea --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_190045/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_190045/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_190045/models_400epochs/tvae_400epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 400)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..e0a553031080d6e3aafd05defa5efd78f7594c9a --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:565273f218660439adfce680de0c5dd47ac78c0168501264f90b50c747ad3d9d +size 248 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/gen_20260520_200013.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/gen_20260520_200013.log new file mode 100644 index 0000000000000000000000000000000000000000..4b82f74eb65568873f9b6a65b8a8d5573f732bab --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/gen_20260520_200013.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d15a7628667c3cd4a25ea1a19c113066abb7c7c192f59170fbc60d5c0156135 +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/models_400epochs/train_20260520_190045.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/models_400epochs/train_20260520_190045.log new file mode 100644 index 0000000000000000000000000000000000000000..efcfb7d93145fd6cbb2e4c16e78f04e3977de3c3 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/models_400epochs/train_20260520_190045.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fa1052bb7c9aa501e211c9b065a1aa07ad541a35f8dfcbbcb131ae63a176eba +size 663 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/models_400epochs/tvae_400epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/models_400epochs/tvae_400epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..dd397fccf3f3f5be245c1375c5bcb589bc23cbb6 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/models_400epochs/tvae_400epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:877617986daaa6939e96f28b44e1e11f8aa6491464cf32a2a69e00289db30b88 +size 5616038 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..aedbd929009abdcdce5bfbb4855f99c724ec8c8c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4715471d3886038835b7499e6304737b9f784b32a3ca6baab859ebd9f4efa84d +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..6da31d67016fea548831d97c0c83e49b49b4054e --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b6175be562fd355d51e0c82336b9e7ce744eff3face01d33752d30153443482 +size 2437 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..4e4a9fdbe2bd3e29fe0a9a1219a1dbb0c5d30ce9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cff0633c7a38602a0c716d23f15a46c4927b75d5ea2138455a1a6ae3bb80698f +size 904 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..d2025d6727d3940f6a4bee1d88d150037bb677a4 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ce2efc80ea1680c912e76ae20ea035325d49a5db34624c12a9bf0867622593b +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0113e2d675de90598ec3b2717113b205dcf8bc16 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79f9b9d19082b8b7c14ba50ee47be2771e184cdbe5768bcbdb6a887129edd437 +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/tvae-n11-15215-20260520_200013.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/tvae-n11-15215-20260520_200013.csv new file mode 100644 index 0000000000000000000000000000000000000000..2c87218ff611a6a4620dcaa9263326d3962f3079 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/tvae-n11-15215-20260520_200013.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de83a563378f95d140d99a74e9913012cda4c485b5bb6fd89611e9f435e01a6f +size 2885061 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_190045/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/_tvae_generate.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/_tvae_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..8df67566d340db2ca62de3bd6556e26fea7c5e76 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/_tvae_generate.py @@ -0,0 +1,27 @@ +import os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +from src.SpecificModels.ctgan_rdt_inverse_fix import apply_ctgan_inverse_fix +apply_ctgan_inverse_fix() +import pandas as pd +from ctgan.synthesizers.tvae import TVAE +def _fit_cpu_staged_compat(self, *args, **kwargs): + raise RuntimeError("_fit_cpu_staged_compat should never be called during generate-only load") +if not hasattr(TVAE, "_fit_cpu_staged"): + TVAE._fit_cpu_staged = _fit_cpu_staged_compat +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +model = TVAE.load("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_200132/models_250epochs/tvae_250epochs.pt") +total = 15215 +chunk = min(50000, total) if total > 50000 else total +parts = [] +left = total +while left > 0: + take = min(chunk, left) + parts.append(model.sample(take)) + left -= take +samples = pd.concat(parts, ignore_index=True) if len(parts) > 1 else parts[0] +samples.to_csv("/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_200132/tvae-n11-15215-20260520_204109.csv", index=False) +print(f"[TVAE] Generated {total} rows (chunks={len(parts)}) -> /work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_200132/tvae-n11-15215-20260520_204109.csv") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/_tvae_train.py b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/_tvae_train.py new file mode 100644 index 0000000000000000000000000000000000000000..615e113312e805e95e60d0b46d167874a24bc173 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/_tvae_train.py @@ -0,0 +1,131 @@ +import json, os, sys +sys.path.insert(0, "/work") +from src.SpecificModels.ctgan_joblib_parallel_cap import apply_parallel_cap_from_env +apply_parallel_cap_from_env() +import pandas as pd +import torch +from torch.optim import Adam +from torch.utils.data import DataLoader, TensorDataset +from tqdm import tqdm +from ctgan.data import read_csv +from ctgan.data_transformer import DataTransformer +from ctgan.synthesizers.tvae import TVAE, Encoder, Decoder, _loss_function + +# Keep transform stage parallelism bounded for stability on shared host. +os.environ.setdefault("LOKY_MAX_CPU_COUNT", "8") +os.environ.setdefault("OPENBLAS_NUM_THREADS", "8") +os.environ.setdefault("MKL_NUM_THREADS", "8") +_nj = (os.environ.get("TVAE_CTGAN_JOBTRANS_N_JOBS") or "").strip() +if _nj: + print("[TVAE] joblib Parallel cap ON, TVAE_CTGAN_JOBTRANS_N_JOBS=" + _nj) +else: + print("[TVAE] joblib Parallel cap OFF (unset TVAE_CTGAN_JOBTRANS_N_JOBS)") +print("[TVAE] LOKY_MAX_CPU_COUNT=" + str(os.environ.get("LOKY_MAX_CPU_COUNT", ""))) + +csv_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_200132/staged/public/train.csv" +meta_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_200132/tvae_metadata.json" +save_path = "/work/output-Benchmark-trainonly-v1/n11/tvae/tvae-n11-20260520_200132/models_250epochs/tvae_250epochs.pt" +epochs = int(os.environ.get("TVAE_EPOCHS", 250)) +batch_size = int(os.environ.get("TVAE_BATCH_SIZE", "500")) +embedding_dim = int(os.environ.get("TVAE_EMBEDDING_DIM", "128")) +l2scale = float(os.environ.get("TVAE_L2SCALE", "1e-5")) +loss_factor = float(os.environ.get("TVAE_LOSS_FACTOR", "2")) +def _parse_dims(name, default): + raw = (os.environ.get(name) or "").strip() + if not raw: + return default + return tuple(int(x.strip()) for x in raw.split(",") if x.strip()) +compress_dims = _parse_dims("TVAE_COMPRESS_DIMS", (128, 128)) +decompress_dims = _parse_dims("TVAE_DECOMPRESS_DIMS", (128, 128)) + +def _fit_cpu_staged(self, train_data, discrete_columns=()): + self.transformer = DataTransformer() + self.transformer.fit(train_data, discrete_columns) + train_data = self.transformer.transform(train_data) + pin_memory = getattr(self._device, "type", str(self._device)) == "cuda" + dataset = TensorDataset(torch.from_numpy(train_data.astype("float32"))) + loader = DataLoader( + dataset, + batch_size=self.batch_size, + shuffle=True, + drop_last=False, + pin_memory=pin_memory, + ) + + data_dim = self.transformer.output_dimensions + encoder = Encoder(data_dim, self.compress_dims, self.embedding_dim).to(self._device) + self.decoder = Decoder(self.embedding_dim, self.decompress_dims, data_dim).to(self._device) + optimizerAE = Adam( + list(encoder.parameters()) + list(self.decoder.parameters()), + weight_decay=self.l2scale, + ) + + self.loss_values = pd.DataFrame(columns=["Epoch", "Batch", "Loss"]) + iterator = tqdm(range(self.epochs), disable=(not self.verbose)) + if self.verbose: + iterator_description = "Loss: {loss:.3f}" + iterator.set_description(iterator_description.format(loss=0)) + + for i in iterator: + loss_values = [] + batch = [] + for id_, data in enumerate(loader): + optimizerAE.zero_grad() + real = data[0].to(self._device, non_blocking=pin_memory) + mu, std, logvar = encoder(real) + eps = torch.randn_like(std) + emb = eps * std + mu + rec, sigmas = self.decoder(emb) + loss_1, loss_2 = _loss_function( + rec, + real, + sigmas, + mu, + logvar, + self.transformer.output_info_list, + self.loss_factor, + ) + loss = loss_1 + loss_2 + loss.backward() + optimizerAE.step() + self.decoder.sigma.data.clamp_(0.01, 1.0) + + batch.append(id_) + loss_values.append(loss.detach().cpu().item()) + + epoch_loss_df = pd.DataFrame({ + "Epoch": [i] * len(batch), + "Batch": batch, + "Loss": loss_values, + }) + if not self.loss_values.empty: + self.loss_values = pd.concat( + [self.loss_values, epoch_loss_df] + ).reset_index(drop=True) + else: + self.loss_values = epoch_loss_df + + if self.verbose: + iterator.set_description( + iterator_description.format(loss=loss.detach().cpu().item()) + ) + +data, discrete_columns = read_csv(csv_path, meta_path, header=True, discrete=None) +print(f"[TVAE] Training on {len(data)} rows, {len(data.columns)} cols, epochs={epochs}, batch_size={batch_size}, embedding_dim={embedding_dim}, compress_dims={compress_dims}, decompress_dims={decompress_dims}, l2scale={l2scale}, loss_factor={loss_factor}") +_orig_fit = TVAE.fit +TVAE.fit = _fit_cpu_staged +try: + model = TVAE( + epochs=epochs, + batch_size=batch_size, + embedding_dim=embedding_dim, + compress_dims=compress_dims, + decompress_dims=decompress_dims, + l2scale=l2scale, + loss_factor=loss_factor, + ) + model.fit(data, discrete_columns) +finally: + TVAE.fit = _orig_fit +model.save(save_path) +print(f"[TVAE] Model saved -> {save_path}") diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/bo_combo.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/bo_combo.json new file mode 100644 index 0000000000000000000000000000000000000000..ae355c506226ce2c5fa6d0564f5aae6ed1e77500 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/bo_combo.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e45340863781aa5a4c427851e4c06587d62638ab465efea48274ec12868afdc +size 261 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/gen_20260520_204109.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/gen_20260520_204109.log new file mode 100644 index 0000000000000000000000000000000000000000..af078988b206328c5f5a88389b29648cdca7aee0 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/gen_20260520_204109.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21912249b237af941e15bc7bd6f1a2d81cc63af59da7dd80ede7330406661aeb +size 409 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/input_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/input_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..06e6c5fd80c3dcb18bc9156e51c573bdee081242 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/input_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608e48ec09d1ed64b35c42a4317e8e79cd812bd576b810d0e45a122bb7beb4cc +size 1357 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/models_250epochs/train_20260520_200132.log b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/models_250epochs/train_20260520_200132.log new file mode 100644 index 0000000000000000000000000000000000000000..de794744c8878a08d236813bb7a17a6a5bc0b69f --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/models_250epochs/train_20260520_200132.log @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89a8a88acce80ac41cb54c94acdf021f088e2eb844a6a1f37dc62d5cdd3a2fb7 +size 667 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/models_250epochs/tvae_250epochs.pt b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/models_250epochs/tvae_250epochs.pt new file mode 100644 index 0000000000000000000000000000000000000000..354997674a29a94b4c483935ebf6957d7ca084e8 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/models_250epochs/tvae_250epochs.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9683bbc61ca60e19d5fc893374717945444db72857016d0b2b520e660341869e +size 3637228 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/public_gate/normalized_schema_snapshot.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/public_gate/normalized_schema_snapshot.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5dd2b4d66e54e5dbeff78b827bc48cfa7dbdf1 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/public_gate/normalized_schema_snapshot.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c3209b66f9891477ccb430a3d02e3a299eb9c1a2d045bbf528bdfeb8de4fb0 +size 5283 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/public_gate/public_gate_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/public_gate/public_gate_report.json new file mode 100644 index 0000000000000000000000000000000000000000..59dc44de856101118ae1288822ef9be24e1daf4c --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/public_gate/public_gate_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6336138aac04ffe8bd9b1596f9d11556e4417d85ddb950dd20c9532ef6b9fd80 +size 915 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/public_gate/staged_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/public_gate/staged_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..45db644183c6fbad10e5c25591fc76a600649bba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/public_gate/staged_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a96f2b620501230c827f01318d8b43d329035c1917c406c5b958b8ecc1e67aa5 +size 6069 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/run_config.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/run_config.json new file mode 100644 index 0000000000000000000000000000000000000000..049549d444c9ad45afd10982e03a3c6452a41584 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/run_config.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecf1023f46a5fbebc5c35b1b5f1677a6c9f3b296eb73d7641f24d83fce4e7137 +size 2441 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/runtime_result.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/runtime_result.json new file mode 100644 index 0000000000000000000000000000000000000000..a8f968e31a69f76b8b1586e74db89fadc1d45f56 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/runtime_result.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1631bf0d8275d5553101042710d6a8d2c76b17b7a0037c576387d230ca0a9197 +size 904 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/staged_features.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/staged_features.json new file mode 100644 index 0000000000000000000000000000000000000000..38380510b7674c7cdd856f51da52bd91fa879aba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/staged_features.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea5e7a5363e366f12e9794c57325c7f28e71ef584464c2f78fa67b80eca9a82 +size 1025 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/test.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/test.csv new file mode 100644 index 0000000000000000000000000000000000000000..59e413eb15b50a2ac49cd95bbe57b0d0018ccf95 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/test.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4e935f84569b849b662718eaa19be605037d46a5f7391f164728d6d9c3bb50 +size 148017 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/train.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/train.csv new file mode 100644 index 0000000000000000000000000000000000000000..cc5acd9fe00e8a233592b4da9797300adc83744b --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/train.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ab621ac8239797506ce800c1409422452b8127da93add91dd9e6d63ddeec6f7 +size 1182326 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/val.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/val.csv new file mode 100644 index 0000000000000000000000000000000000000000..e0a3b6ab9c53a7f8d97dff3290e643431f9ad023 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/public/val.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374e44edf25e58ca37eabbef0485c98cecdbdfa9d24da738937f404840bfad44 +size 147784 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/tvae/adapter_report.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/tvae/adapter_report.json new file mode 100644 index 0000000000000000000000000000000000000000..481c12efa9afdd3af1e2a3c766f18f1c1655c54d --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/tvae/adapter_report.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a7da07a9468f853c3112b08df1ebd81ea0319ab3cf021bdb313afc48815608f +size 314 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/tvae/adapter_transforms_applied.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/tvae/adapter_transforms_applied.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3aaa01a627a955d2022165805d2786c0da78ba --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/tvae/adapter_transforms_applied.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 +size 2 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/tvae/model_input_manifest.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/tvae/model_input_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9acd3c5cf9efb38907d7ba76632e2c7a625cdb02 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/staged/tvae/model_input_manifest.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efa3d5e21557dc9de72a5f1ca93232ef22ed5aae4ea2fdc48b6253d7dd7b2ffd +size 6259 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/tvae-n11-15215-20260520_204109.csv b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/tvae-n11-15215-20260520_204109.csv new file mode 100644 index 0000000000000000000000000000000000000000..af8279c50ab4785c22f842899445c99079499723 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/tvae-n11-15215-20260520_204109.csv @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91a880a4e10620e2dd6e37e37b9c9d5cfd7f11ef0c0170e7f99c5e6e2ae883ff +size 2884646 diff --git a/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/tvae_metadata.json b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/tvae_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..67a0338db890724baf86effd9cce456b4bfcdeb9 --- /dev/null +++ b/Hyper ParameterTuning/n11/tvae/runs/tvae-n11-20260520_200132/tvae_metadata.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600119beca783dcd77c3bd4873a4dc3b4c207ab14bb1fceeb482b78deec4e45d +size 726