Add personalised track models (frozen backbones + PRA adapters, offline/online x 3 conditions)

personalised/code/configs/README.md

@@ -0,0 +1,27 @@
+# Config Organization
+
+The task-facing Hydra configs are organized by challenge setting. Use these files with `python main.py --config-name <section>/<method> ...`.
+
+| Section | Entry config | Training / evaluation entry | Model config files | Model code | Checkpoints |
+| --- | --- | --- | --- | --- | --- |
+| Generic online | `generic_online/motion_diffusion.yaml` | `main.py` with `stage=fit` or `stage=test` | `generic_online/model/motion_diffusion.yaml`, `generic_online/model/losses/motion_diffusion.yaml` | `trainer/motion_diffusion.py`, `framework/motion_diffusion/` | `save/motion_diffusion/react_2025/online/checkpoints/<run_id>/` |
+| Generic online | `generic_online/motion_transvae.yaml` | `main.py` with `stage=fit` or `stage=test` | `generic_online/model/motion_transvae.yaml`, `generic_online/model/losses/motion_transvae.yaml` | `trainer/motion_transvae.py`, `framework/motion_transvae/` | `save/motion_transvae/react_2025/online/checkpoints/<run_id>/` |
+| Generic offline | `generic_offline/motion_diffusion.yaml` | `main.py` with `stage=fit` or `stage=test` | `generic_offline/model/motion_diffusion.yaml`, `generic_offline/model/losses/motion_diffusion.yaml` | `trainer/motion_diffusion.py`, `framework/motion_diffusion/` | `save/motion_diffusion/react_2025/offline/checkpoints/<run_id>/` |
+| Generic offline | `generic_offline/motion_transvae.yaml` | `main.py` with `stage=fit` or `stage=test` | `generic_offline/model/motion_transvae.yaml`, `generic_offline/model/losses/motion_transvae.yaml` | `trainer/motion_transvae.py`, `framework/motion_transvae/` | `save/motion_transvae/react_2025/offline/checkpoints/<run_id>/` |
+| Generic offline | RegNN | `regnn/train.py`; add `--test` for evaluation | RegNN uses command-line arguments instead of Hydra YAML | `regnn/models/`, `regnn/trainers.py` | `regnn/<logs-dir>/mhp-*-seed<seed>.pth` or `regnn/<logs-dir>/mhp-eeg-head-*-seed<seed>.pth` |
+| Personalized online | `personalized_online/perfrdiff_rewrite_weight.yaml` | `main.py` with `stage=fit` or `stage=test` | `personalized_online/model/motion_diffusion.yaml`, `personalized_online/model/losses/perfrdiff_rewrite_weight.yaml` | `trainer/perfrdiff_rewrite_weight.py`, `framework/perfrdiff_rewrite_weight/` | `save/perfrdiff_rewrite_weight/react_2025/online/checkpoints/<run_id>/` |
+| Personalized offline | `personalized_offline/perfrdiff_rewrite_weight.yaml` | `main.py` with `stage=fit` or `stage=test` | `personalized_offline/model/motion_diffusion.yaml`, `personalized_offline/model/losses/perfrdiff_rewrite_weight.yaml` | `trainer/perfrdiff_rewrite_weight.py`, `framework/perfrdiff_rewrite_weight/` | `save/perfrdiff_rewrite_weight/react_2025/offline/checkpoints/<run_id>/` |
+
+## Directory Notes
+
+- `data/`, `trainer/`, `model/`, and `model/losses/` under each section contain the task-specific YAML used by that section.
+- `shared/` contains global support configs: `data/`, `logger/`, `model/`, `trainer/`, and `path.yaml`.
+- `configs/shared/path.yaml` exposes resolver-based paths such as the code root and current working directory, so it is kept with the other shared configs.
+- `configs/shared/model/emotion_autoencoder.yaml` is intentionally kept as a shared config because the post-processor loads it by this path.
+
+## Running
+
+- Hydra training: `python main.py --config-name <section>/<method> stage=fit data_dir=<dataset-root>`
+- Hydra evaluation: `python main.py --config-name <section>/<method> stage=test data_dir=<dataset-root> resume_id=<run-id>`
+- RegNN training: run `python train.py ...` from the `regnn/` directory.
+- RegNN evaluation: run `python train.py --test ...` from the `regnn/` directory.

personalised/code/configs/data/emotion_autoencoder.yaml

@@ -0,0 +1,29 @@
+data_name: react_2025
+_target_: dataset.react_2025.ReactionDataloader
+
+train_dataset:
+    _target_: dataset.react_2025.ReactionAutoEncoderDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: True
+    num_workers: 2
+    root_dir: ${data_dir}
+    split: train
+    clip_length: ${trainer.clip_length}
+
+validation_dataset:
+    _target_: dataset.react_2025.ReactionAutoEncoderDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 2
+    root_dir: ${data_dir}
+    split: val
+    clip_length: ${trainer.clip_length}
+
+test_dataset:
+    _target_: dataset.react_2025.ReactionAutoEncoderDataset
+    batch_size: 1
+    shuffle: False
+    num_workers: 2
+    root_dir: ${data_dir}
+    split: test
+    clip_length: ${trainer.clip_length}

personalised/code/configs/data/motion_diffusion.yaml

@@ -0,0 +1,122 @@
+data_name: react_2025
+_target_: dataset.react_2025.ReactionDataloader
+
+train_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: True
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: train
+    clip_length: 750  # 256 (transvae)
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+
+validation_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: val
+    target_size: 224
+    crop_size: 224
+    clip_length: 750  # 256 (transvae)
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+
+test_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 4
+    root_dir: ${data_dir}
+    split: test
+    target_size: 224
+    crop_size: 224
+    clip_length: 750  # 256 (transvae)
+    fps: 30
+    audio_feature_type: wav2vec  # mfcc | wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: ${trainer.generic.eval_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}

personalised/code/configs/data/motion_transvae.yaml

@@ -0,0 +1,125 @@
+data_name: react_2025
+_target_: dataset.react_2025.ReactionDataloader
+
+train_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: True
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: train
+    clip_length: ${trainer.max_seq_len}
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: True
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.bidirectional}
+    normalize_3dmm: ${trainer.data_transform}
+
+validation_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: val
+    target_size: 224
+    crop_size: 224
+    clip_length: ${trainer.max_seq_len}
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: True
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.bidirectional}
+    normalize_3dmm: ${trainer.data_transform}
+
+test_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: 1
+    shuffle: True
+    num_workers: 0
+    root_dir: ${data_dir}
+    split: test
+    target_size: 224
+    crop_size: 224
+    clip_length: ${trainer.max_seq_len}
+    fps: 30
+    audio_feature_type: wav2vec  # mfcc | wav2vec
+    load_video_s: True
+    load_video_l: ${trainer.renderer.do_render}
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: ${trainer.eval_facial_metrics}
+    load_3dmm_s: True
+    load_3dmm_l: False
+    load_eeg_l: ${trainer.eval_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.bidirectional}
+    normalize_3dmm: ${trainer.data_transform}

personalised/code/configs/data/perfrdiff_rewrite_weight.yaml

@@ -0,0 +1,135 @@
+data_name: react_2025
+_target_: dataset.perfrdiff_rewrite_weight.PerFRDiffRewriteWeightDataModule
+
+train_dataset:
+    _target_: dataset.perfrdiff_rewrite_weight.PerFRDiffRewriteWeightDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: True
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: train
+    clip_length: 750
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_ref: True
+    load_personality_l: True
+    personal_condition_mode: ${trainer.main_model.args.personal_condition_mode}
+    load_eeg_l: ${trainer.generic.train_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+    normalize_3dmm: standard
+
+validation_dataset:
+    _target_: dataset.perfrdiff_rewrite_weight.PerFRDiffRewriteWeightDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: val
+    clip_length: 750
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_ref: True
+    load_personality_l: True
+    personal_condition_mode: ${trainer.main_model.args.personal_condition_mode}
+    load_eeg_l: ${trainer.generic.train_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+    normalize_3dmm: standard
+
+test_dataset:
+    _target_: dataset.perfrdiff_rewrite_weight.PerFRDiffRewriteWeightDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 4
+    root_dir: ${data_dir}
+    split: test
+    clip_length: 750
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_ref: False
+    load_personality_l: True
+    personal_condition_mode: ${trainer.main_model.args.personal_condition_mode}
+    load_eeg_l: ${trainer.generic.eval_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+    normalize_3dmm: standard
+    num_test_gts: ${num_gts}

personalised/code/configs/g2p_delta.yaml

@@ -0,0 +1,88 @@
+defaults:
+  - /personalized_offline/data/perfrdiff_rewrite_weight@data
+  - /personalized_offline/trainer/perfrdiff_rewrite_weight@trainer
+  - /shared/logger: none
+  - /shared/path@path
+  - _self_
+
+seed: 1234
+logger_level: INFO
+task: offline
+stage: ???
+data_dir: ???
+resume: false
+run_id: ${generate_id:}
+resume_id: ""
+num_gts: 10
+
+trainer:
+  _target_: trainer.g2p_delta.Trainer
+  model_name: g2p_delta
+  batch_size: 2
+  pretrained:
+    diffusion_decoder: pretrained_models/generic_offline_final_submission_model/CausalTransformerDenoiser/checkpoint_120.pth
+    diffusion_prior: pretrained_models/generic_offline_final_submission_model/DiffusionPriorNetwork/checkpoint_120.pth
+    modifier_checkpoint: ""
+    eeg_head_checkpoint: pretrained_models/generic_offline_final_submission_model/EEGPredictionHead/checkpoint_120.pth
+  model:
+    diff_model:
+      _target_: framework.motion_diffusion.diffusion.matchers_causal.CausalLatentMatcher
+      eeg_head:
+        enabled: true
+      diffusion_decoder:
+        args:
+          use_lag_bias: true
+          lag_max: 60
+          lag_lookahead: 0
+          use_coarse: true
+          coarse_classes: 8
+          coarse_hidden: 256
+          coarse_emo_start: 17
+  criterion:
+    args:
+      eeg_loss_weight: 0.25
+  generic:
+    epochs: 30
+    save_period: 5
+    val_period: 5
+    clip_grad: true
+    train_eeg: true
+    train_eeg_head_only: false
+    eval_eeg: true
+    num_workers: 8
+    num_preds: 1
+    max_train_batches: 0
+    max_val_batches: 0
+    # Short-triage control only: matched (default) | shuffled | identity.
+    # See trainer/g2p_delta.py::_apply_personalization and
+    # scripts/build_subset_eval_root.py. Never used during training.
+    eval_condition_mode: matched
+  main_model:
+    args:
+      personal_condition_mode: personality_only
+      embed_dim: 256
+      history_hidden_dim: 128
+      personality_hidden_dim: 128
+      condition_dropout: 0.1
+      delta_rank: 128
+      delta_max_scale: 0.15
+      personalize_coarse: true
+      anchor_weight: 0.001
+      coarse_weight: 0.5
+      ccc_weight: 0.1
+      dynamics_weight: 0.02
+      counterfactual_weight: 0.1
+      counterfactual_margin: 0.02
+      eeg_lr: 0.00001
+    optimizer_hypernet:
+      type: adamw
+      args:
+        lr: 0.0002
+        weight_decay: 0.0001
+        momentum: 0.0
+
+hydra:
+  run:
+    dir: outputs/${trainer.model_name}/${data.data_name}/${task}/${run_id}
+  job:
+    chdir: true

personalised/code/configs/generic_offline/data/motion_diffusion.yaml

@@ -0,0 +1,122 @@
+data_name: react_2025
+_target_: dataset.react_2025.ReactionDataloader
+
+train_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: True
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: train
+    clip_length: 750  # 256 (transvae)
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+
+validation_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: val
+    target_size: 224
+    crop_size: 224
+    clip_length: 750  # 256 (transvae)
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+
+test_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 4
+    root_dir: ${data_dir}
+    split: test
+    target_size: 224
+    crop_size: 224
+    clip_length: 750  # 256 (transvae)
+    fps: 30
+    audio_feature_type: wav2vec  # mfcc | wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: ${trainer.generic.eval_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}

personalised/code/configs/generic_offline/data/motion_transvae.yaml

@@ -0,0 +1,125 @@
+data_name: react_2025
+_target_: dataset.react_2025.ReactionDataloader
+
+train_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: True
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: train
+    clip_length: ${trainer.max_seq_len}
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: True
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.bidirectional}
+    normalize_3dmm: ${trainer.data_transform}
+
+validation_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: val
+    target_size: 224
+    crop_size: 224
+    clip_length: ${trainer.max_seq_len}
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: True
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.bidirectional}
+    normalize_3dmm: ${trainer.data_transform}
+
+test_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: 1
+    shuffle: True
+    num_workers: 0
+    root_dir: ${data_dir}
+    split: test
+    target_size: 224
+    crop_size: 224
+    clip_length: ${trainer.max_seq_len}
+    fps: 30
+    audio_feature_type: wav2vec  # mfcc | wav2vec
+    load_video_s: True
+    load_video_l: ${trainer.renderer.do_render}
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: ${trainer.eval_facial_metrics}
+    load_3dmm_s: True
+    load_3dmm_l: False
+    load_eeg_l: ${trainer.eval_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.bidirectional}
+    normalize_3dmm: ${trainer.data_transform}

personalised/code/configs/generic_offline/model/losses/motion_diffusion.yaml

@@ -0,0 +1,11 @@
+_target_: framework.utils.losses.DiffusionLoss
+
+losses_type: MSELossApt
+losses_multiplier: 1.0
+n_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+temporal_loss_w: 0.0
+prior_loss_weight: 1.0
+eeg_loss_weight: 0.25
+w_au: 1.0  # action unit
+w_va: 5.0  # valence and arousal
+w_em: 5.0  # emotion

personalised/code/configs/generic_offline/model/losses/motion_transvae.yaml

@@ -0,0 +1,8 @@
+_target_: framework.utils.losses.VAELoss
+
+kl_p: 0.00001  # kl div loss
+w_emo: 2.0
+w_exp: 2.0  # facial expression
+w_rot: 4.0  # rotation
+w_tran: 4.0  # translation
+eeg_loss_weight: 0.25

personalised/code/configs/generic_offline/model/motion_diffusion.yaml

@@ -0,0 +1,116 @@
+diff_model:
+    model_name: LatentMLPMatcher
+    _target_: framework.motion_diffusion.diffusion.matchers.LatentMatcher
+
+    eeg_head:
+        enabled: True
+        output_dim: 14
+        hidden_dim: 256
+        dropout: 0.5
+        pooling: mean
+        detach_prediction_emotion: True
+        use_speaker_audio: True
+        use_speaker_emotion: True
+        use_speaker_3dmm: True
+        use_prediction_emotion: True
+
+    diffusion_prior:
+        type:
+            DiffusionPriorNetwork
+        args:
+            emb_preprocessing: normalize
+            freeze_encoder: True
+            audio_dim: 768
+            window_size: ${trainer.${trainer.trainer_mode}.window_size}
+            s_ratio: ${trainer.${trainer.trainer_mode}.s_ratio}
+            token_len: 750
+            _3dmm_dim: 58
+            speaker_emb_dim: 512
+            latent_dim: 512
+            depth: 4
+            num_time_layers: 2
+            num_time_embeds: 1
+            num_time_emb_channels: 64
+            time_last_act: False
+            use_learned_query: True
+            s_audio_cond_drop_prob: 0.2
+            s_latentemb_cond_drop_prob: 1.0
+            s_3dmm_cond_drop_prob: 0.2
+            guidance_scale: 1.0
+            dim_head: 64
+            heads: 8
+            ff_mult: 4
+            norm_in: False
+            norm_out: True
+            attn_dropout: 0.0
+            ff_dropout: 0.0
+            final_proj: True
+            normformer: False
+            rotary_emb: True
+        scheduler:
+            noise_schedule: cosine
+            timestep_spacing: leading
+            num_train_timesteps: 1000
+            num_inference_timesteps: 50
+            predict: start_x
+            var_type: fixed_large
+            rescale_timesteps: False
+            noise_std: 1
+            num_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+
+    diffusion_decoder:
+        type:
+            TransformerDenoiser
+        args:
+            emb_preprocessing: normalize
+            freeze_encoder: True
+            window_size: ${trainer.${trainer.trainer_mode}.window_size}
+            s_ratio: ${trainer.${trainer.trainer_mode}.s_ratio}
+            token_len: 750
+            encode_emotion: False
+            encode_3dmm: False
+            ablation_skip_connection: True
+            nfeats: 25
+            latent_dim: 512
+            ff_size: 1024
+            num_layers: 9  # 7 | 9
+            num_heads: 8  # 4 | 8
+            dropout: 0.1
+            normalize_before: False
+            activation: gelu
+            flip_sin_to_cos: True
+            return_intermediate_dec: False
+            position_embedding: learned
+            arch: trans_dec
+            freq_shift: 0
+            time_encoded_dim: 64
+            s_audio_dim: 768
+            s_audio_scale: 1.0
+            s_emotion_dim: 25
+            l_embed_dim: 512
+            s_embed_dim: 512
+            personal_emb_dim: 512
+            s_3dmm_dim: 58
+            concat: concat_first
+            guidance_scale: 1.0  # 7.5
+            s_audio_enc_drop_prob: 0.2
+            s_latent_embed_drop_prob: 1.0
+            s_3dmm_enc_drop_prob: 0.2
+            s_emotion_enc_drop_prob: 0.2
+            past_l_emotion_drop_prob: 1.0
+            use_past_frames: False
+        scheduler:
+            noise_schedule: cosine
+            timestep_spacing: leading
+            num_train_timesteps: 1000
+            num_inference_timesteps: 50
+            predict: start_x
+            var_type: fixed_large
+            rescale_timesteps: False
+            noise_std: 1
+            num_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+
+defaults:
+    - /generic_offline/model/losses/motion_diffusion@loss
+    - /generic_offline/model/motion_diffusion/audio_embedder@audio_encoder
+    #- /model/latent_embedder@latent_embedder

personalised/code/configs/generic_offline/model/motion_diffusion/audio_embedder.yaml

@@ -0,0 +1,4 @@
+type: AudioEmbedder
+_target_: dataset.modules.audio_embedder.AudioEmbedder
+checkpoint_path: ""
+skip_norm: True

personalised/code/configs/generic_offline/model/motion_transvae.yaml

@@ -0,0 +1,24 @@
+model_name: TransformerVAE
+_target_: framework.motion_transvae.TransformerVAE.TransformerVAE
+
+img_size: 224
+audio_dim: 768
+output_3dmm_dim: 58
+output_emotion_dim: 25
+feature_dim: 128
+seq_len: ${data.train_dataset.clip_length}
+task: ${task}
+window_size: ${trainer.window_size}
+device: cuda
+
+eeg_head:
+    enabled: True
+    output_dim: 14
+    hidden_dim: 256
+    dropout: 0.5
+    pooling: mean
+    detach_prediction_emotion: True
+    use_speaker_audio: True
+    use_speaker_emotion: True
+    use_speaker_3dmm: True
+    use_prediction_emotion: True

personalised/code/configs/generic_offline/motion_diffusion.yaml

@@ -0,0 +1,23 @@
+hydra:
+  run:
+    dir: outputs/${trainer.model_name}/${data.data_name}/${task}/${run_id}
+  job:
+    chdir: true
+
+seed: 1234
+logger_level: INFO
+
+task: offline
+stage: ???
+data_dir: ???
+resume: false
+run_id: ${generate_id:}
+resume_id: ""
+num_gts: 10
+
+defaults:
+  - /generic_offline/data/motion_diffusion@data
+  - /generic_offline/trainer/motion_diffusion@trainer
+  - /shared/logger: none
+  - /shared/path@path
+  - _self_

personalised/code/configs/generic_offline/motion_transvae.yaml

@@ -0,0 +1,23 @@
+hydra:
+  run:
+    dir: outputs/${trainer.model_name}/${data.data_name}/${task}/${run_id}
+  job:
+    chdir: true
+
+seed: 1234
+logger_level: INFO
+
+task: offline
+stage: ???
+data_dir: ???
+resume: false
+run_id: ${generate_id:}
+resume_id: ""
+num_gts: 10
+
+defaults:
+  - /generic_offline/data/motion_transvae@data
+  - /generic_offline/trainer/motion_transvae@trainer
+  - /shared/logger: none
+  - /shared/path@path
+  - _self_

personalised/code/configs/generic_offline/trainer/motion_diffusion.yaml

@@ -0,0 +1,48 @@
+model_name: motion_diffusion
+_target_: trainer.motion_diffusion.Trainer
+
+trainer_mode: generic  # generic facial reaction generator
+folder: save/${trainer.model_name}/${data.data_name}/${task}
+ckpt_dir: ${get_last_checkpoint:${trainer.folder}}
+resumed_training: ${resume}  # true | false
+current_runid: ${run_id}
+resume_runid: ${resume_id}
+batch_size: 4
+post_config_name: configs/shared/model/emotion_autoencoder.yaml
+post_clip_length: 1000  # ground-truths alignment
+data_clamp: True
+
+generic:
+    seed: 1234
+    start_epoch: 0
+    epochs: 500
+    lr: 0.0001
+    model: LatentMatcher
+    clip_grad: False
+    num_workers: 16
+    log_dir: log
+    tb_dir: tb_logs
+    out_dir: results
+    save_period: 10
+    val_period: 10
+    num_preds: 10
+    clip_length: 750
+    window_size: 30
+    s_ratio: 2
+    train_eeg_head_only: False
+    pretrained_decoder_checkpoint: ""
+    pretrained_prior_checkpoint: ""
+    pretrained_load_prior: True
+    eval_eeg: True
+    bidirectional: False
+
+renderer:  # face animation based on predicted 3DMM coefficients
+    name: PIRenderer
+    _target_: utils.render.Render
+    do_render: False
+
+defaults:
+    - /generic_offline/model/motion_diffusion@model
+    - /shared/model/optim/adamw@optim
+    - /shared/model/scheduler/cosine_annealing@scheduler
+    - /generic_offline/model/losses/motion_diffusion@criterion  # loss func

personalised/code/configs/generic_offline/trainer/motion_transvae.yaml

@@ -0,0 +1,42 @@
+model_name: motion_transvae
+_target_: trainer.motion_transvae.Trainer
+
+folder: save/${trainer.model_name}/${data.data_name}/${task}
+ckpt_dir: ${get_last_checkpoint:${trainer.folder}}
+resumed_training: ${resume}  # true | false
+current_runid: ${run_id}
+resume_runid: ${resume_id}
+
+epochs: 100
+gpu_ids: 0
+lr: 0.00001
+j: 12  # num_workers
+batch_size: 4
+max_seq_len: 750
+window_size: 8
+div_p: 10
+post_config_name: configs/shared/model/emotion_autoencoder.yaml
+post_clip_length: 1000  # ground-truths alignment
+data_clamp: True
+data_transform: zero_center  # standard | zero_center
+train_eeg_head_only: False
+pretrained_model_checkpoint: ""
+eval_eeg: True
+num_preds: 10
+bidirectional: False
+save_results: True
+eval_facial_metrics: True
+eval_eeg_metrics: True
+metric_threads: 1
+eval_clip_batch_size: 1
+
+renderer:  # face animation based on predicted 3DMM coefficients
+    name: PIRenderer
+    _target_: utils.render.Render
+    do_render: True
+    transform_reverse: ${trainer.data_transform}
+
+defaults:
+    - /generic_offline/model/motion_transvae@model
+    - /shared/model/optim/adamw@optim
+    - /generic_offline/model/losses/motion_transvae@criterion  # loss func

personalised/code/configs/generic_online/data/motion_diffusion.yaml

@@ -0,0 +1,122 @@
+data_name: react_2025
+_target_: dataset.react_2025.ReactionDataloader
+
+train_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: True
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: train
+    clip_length: 750  # 256 (transvae)
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+
+validation_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: val
+    target_size: 224
+    crop_size: 224
+    clip_length: 750  # 256 (transvae)
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+
+test_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 4
+    root_dir: ${data_dir}
+    split: test
+    target_size: 224
+    crop_size: 224
+    clip_length: 750  # 256 (transvae)
+    fps: 30
+    audio_feature_type: wav2vec  # mfcc | wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: ${trainer.generic.eval_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}

personalised/code/configs/generic_online/data/motion_transvae.yaml

@@ -0,0 +1,125 @@
+data_name: react_2025
+_target_: dataset.react_2025.ReactionDataloader
+
+train_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: True
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: train
+    clip_length: ${trainer.max_seq_len}
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: True
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.bidirectional}
+    normalize_3dmm: ${trainer.data_transform}
+
+validation_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: val
+    target_size: 224
+    crop_size: 224
+    clip_length: ${trainer.max_seq_len}
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: True
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_eeg_l: True
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.bidirectional}
+    normalize_3dmm: ${trainer.data_transform}
+
+test_dataset:
+    _target_: dataset.react_2025.ReactionDataset
+    batch_size: 1
+    shuffle: True
+    num_workers: 0
+    root_dir: ${data_dir}
+    split: test
+    target_size: 224
+    crop_size: 224
+    clip_length: ${trainer.max_seq_len}
+    fps: 30
+    audio_feature_type: wav2vec  # mfcc | wav2vec
+    load_video_s: True
+    load_video_l: ${trainer.renderer.do_render}
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: ${trainer.eval_facial_metrics}
+    load_3dmm_s: True
+    load_3dmm_l: False
+    load_eeg_l: ${trainer.eval_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.bidirectional}
+    normalize_3dmm: ${trainer.data_transform}

personalised/code/configs/generic_online/model/losses/motion_diffusion.yaml

@@ -0,0 +1,11 @@
+_target_: framework.utils.losses.DiffusionLoss
+
+losses_type: MSELossApt
+losses_multiplier: 1.0
+n_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+temporal_loss_w: 0.0
+prior_loss_weight: 1.0
+eeg_loss_weight: 0.25
+w_au: 1.0  # action unit
+w_va: 5.0  # valence and arousal
+w_em: 5.0  # emotion

personalised/code/configs/generic_online/model/losses/motion_transvae.yaml

@@ -0,0 +1,8 @@
+_target_: framework.utils.losses.VAELoss
+
+kl_p: 0.00001  # kl div loss
+w_emo: 2.0
+w_exp: 2.0  # facial expression
+w_rot: 4.0  # rotation
+w_tran: 4.0  # translation
+eeg_loss_weight: 0.25

personalised/code/configs/generic_online/model/motion_diffusion.yaml

@@ -0,0 +1,116 @@
+diff_model:
+    model_name: LatentMLPMatcher
+    _target_: framework.motion_diffusion.diffusion.matchers.LatentMatcher
+
+    eeg_head:
+        enabled: True
+        output_dim: 14
+        hidden_dim: 256
+        dropout: 0.5
+        pooling: mean
+        detach_prediction_emotion: True
+        use_speaker_audio: True
+        use_speaker_emotion: True
+        use_speaker_3dmm: True
+        use_prediction_emotion: True
+
+    diffusion_prior:
+        type:
+            DiffusionPriorNetwork
+        args:
+            emb_preprocessing: normalize
+            freeze_encoder: True
+            audio_dim: 768
+            window_size: ${trainer.${trainer.trainer_mode}.window_size}
+            s_ratio: ${trainer.${trainer.trainer_mode}.s_ratio}
+            token_len: 750
+            _3dmm_dim: 58
+            speaker_emb_dim: 512
+            latent_dim: 512
+            depth: 4
+            num_time_layers: 2
+            num_time_embeds: 1
+            num_time_emb_channels: 64
+            time_last_act: False
+            use_learned_query: True
+            s_audio_cond_drop_prob: 0.2
+            s_latentemb_cond_drop_prob: 1.0
+            s_3dmm_cond_drop_prob: 0.2
+            guidance_scale: 1.0
+            dim_head: 64
+            heads: 8
+            ff_mult: 4
+            norm_in: False
+            norm_out: True
+            attn_dropout: 0.0
+            ff_dropout: 0.0
+            final_proj: True
+            normformer: False
+            rotary_emb: True
+        scheduler:
+            noise_schedule: cosine
+            timestep_spacing: leading
+            num_train_timesteps: 1000
+            num_inference_timesteps: 50
+            predict: start_x
+            var_type: fixed_large
+            rescale_timesteps: False
+            noise_std: 1
+            num_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+
+    diffusion_decoder:
+        type:
+            TransformerDenoiser
+        args:
+            emb_preprocessing: normalize
+            freeze_encoder: True
+            window_size: ${trainer.${trainer.trainer_mode}.window_size}
+            s_ratio: ${trainer.${trainer.trainer_mode}.s_ratio}
+            token_len: 750
+            encode_emotion: False
+            encode_3dmm: False
+            ablation_skip_connection: True
+            nfeats: 25
+            latent_dim: 512
+            ff_size: 1024
+            num_layers: 9  # 7 | 9
+            num_heads: 8  # 4 | 8
+            dropout: 0.1
+            normalize_before: False
+            activation: gelu
+            flip_sin_to_cos: True
+            return_intermediate_dec: False
+            position_embedding: learned
+            arch: trans_dec
+            freq_shift: 0
+            time_encoded_dim: 64
+            s_audio_dim: 768
+            s_audio_scale: 1.0
+            s_emotion_dim: 25
+            l_embed_dim: 512
+            s_embed_dim: 512
+            personal_emb_dim: 512
+            s_3dmm_dim: 58
+            concat: concat_first
+            guidance_scale: 1.0  # 7.5
+            s_audio_enc_drop_prob: 0.2
+            s_latent_embed_drop_prob: 1.0
+            s_3dmm_enc_drop_prob: 0.2
+            s_emotion_enc_drop_prob: 0.2
+            past_l_emotion_drop_prob: 1.0
+            use_past_frames: False
+        scheduler:
+            noise_schedule: cosine
+            timestep_spacing: leading
+            num_train_timesteps: 1000
+            num_inference_timesteps: 50
+            predict: start_x
+            var_type: fixed_large
+            rescale_timesteps: False
+            noise_std: 1
+            num_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+
+defaults:
+    - /generic_online/model/losses/motion_diffusion@loss
+    - /generic_online/model/motion_diffusion/audio_embedder@audio_encoder
+    #- /model/latent_embedder@latent_embedder

personalised/code/configs/generic_online/model/motion_diffusion/audio_embedder.yaml

@@ -0,0 +1,4 @@
+type: AudioEmbedder
+_target_: dataset.modules.audio_embedder.AudioEmbedder
+checkpoint_path: ""
+skip_norm: True

personalised/code/configs/generic_online/model/motion_transvae.yaml

@@ -0,0 +1,24 @@
+model_name: TransformerVAE
+_target_: framework.motion_transvae.TransformerVAE.TransformerVAE
+
+img_size: 224
+audio_dim: 768
+output_3dmm_dim: 58
+output_emotion_dim: 25
+feature_dim: 128
+seq_len: ${data.train_dataset.clip_length}
+task: ${task}
+window_size: ${trainer.window_size}
+device: cuda
+
+eeg_head:
+    enabled: True
+    output_dim: 14
+    hidden_dim: 256
+    dropout: 0.5
+    pooling: mean
+    detach_prediction_emotion: True
+    use_speaker_audio: True
+    use_speaker_emotion: True
+    use_speaker_3dmm: True
+    use_prediction_emotion: True

personalised/code/configs/generic_online/motion_diffusion.yaml

@@ -0,0 +1,23 @@
+hydra:
+  run:
+    dir: outputs/${trainer.model_name}/${data.data_name}/${task}/${run_id}
+  job:
+    chdir: true
+
+seed: 1234
+logger_level: INFO
+
+task: online
+stage: ???
+data_dir: ???
+resume: false
+run_id: ${generate_id:}
+resume_id: ""
+num_gts: 10
+
+defaults:
+  - /generic_online/data/motion_diffusion@data
+  - /generic_online/trainer/motion_diffusion@trainer
+  - /shared/logger: none
+  - /shared/path@path
+  - _self_

personalised/code/configs/generic_online/motion_transvae.yaml

@@ -0,0 +1,23 @@
+hydra:
+  run:
+    dir: outputs/${trainer.model_name}/${data.data_name}/${task}/${run_id}
+  job:
+    chdir: true
+
+seed: 1234
+logger_level: INFO
+
+task: online
+stage: ???
+data_dir: ???
+resume: false
+run_id: ${generate_id:}
+resume_id: ""
+num_gts: 10
+
+defaults:
+  - /generic_online/data/motion_transvae@data
+  - /generic_online/trainer/motion_transvae@trainer
+  - /shared/logger: none
+  - /shared/path@path
+  - _self_

personalised/code/configs/generic_online/trainer/motion_diffusion.yaml

@@ -0,0 +1,48 @@
+model_name: motion_diffusion
+_target_: trainer.motion_diffusion.Trainer
+
+trainer_mode: generic  # generic facial reaction generator
+folder: save/${trainer.model_name}/${data.data_name}/${task}
+ckpt_dir: ${get_last_checkpoint:${trainer.folder}}
+resumed_training: ${resume}  # true | false
+current_runid: ${run_id}
+resume_runid: ${resume_id}
+batch_size: 4
+post_config_name: configs/shared/model/emotion_autoencoder.yaml
+post_clip_length: 1000  # ground-truths alignment
+data_clamp: True
+
+generic:
+    seed: 1234
+    start_epoch: 0
+    epochs: 500
+    lr: 0.0001
+    model: LatentMatcher
+    clip_grad: False
+    num_workers: 16
+    log_dir: log
+    tb_dir: tb_logs
+    out_dir: results
+    save_period: 10
+    val_period: 10
+    num_preds: 10
+    clip_length: 750
+    window_size: 30
+    s_ratio: 2
+    train_eeg_head_only: False
+    pretrained_decoder_checkpoint: ""
+    pretrained_prior_checkpoint: ""
+    pretrained_load_prior: True
+    eval_eeg: True
+    bidirectional: False
+
+renderer:  # face animation based on predicted 3DMM coefficients
+    name: PIRenderer
+    _target_: utils.render.Render
+    do_render: False
+
+defaults:
+    - /generic_online/model/motion_diffusion@model
+    - /shared/model/optim/adamw@optim
+    - /shared/model/scheduler/cosine_annealing@scheduler
+    - /generic_online/model/losses/motion_diffusion@criterion  # loss func

personalised/code/configs/generic_online/trainer/motion_transvae.yaml

@@ -0,0 +1,42 @@
+model_name: motion_transvae
+_target_: trainer.motion_transvae.Trainer
+
+folder: save/${trainer.model_name}/${data.data_name}/${task}
+ckpt_dir: ${get_last_checkpoint:${trainer.folder}}
+resumed_training: ${resume}  # true | false
+current_runid: ${run_id}
+resume_runid: ${resume_id}
+
+epochs: 100
+gpu_ids: 0
+lr: 0.00001
+j: 12  # num_workers
+batch_size: 4
+max_seq_len: 750
+window_size: 8
+div_p: 10
+post_config_name: configs/shared/model/emotion_autoencoder.yaml
+post_clip_length: 1000  # ground-truths alignment
+data_clamp: True
+data_transform: zero_center  # standard | zero_center
+train_eeg_head_only: False
+pretrained_model_checkpoint: ""
+eval_eeg: True
+num_preds: 10
+bidirectional: False
+save_results: True
+eval_facial_metrics: True
+eval_eeg_metrics: True
+metric_threads: 1
+eval_clip_batch_size: 1
+
+renderer:  # face animation based on predicted 3DMM coefficients
+    name: PIRenderer
+    _target_: utils.render.Render
+    do_render: True
+    transform_reverse: ${trainer.data_transform}
+
+defaults:
+    - /generic_online/model/motion_transvae@model
+    - /shared/model/optim/adamw@optim
+    - /generic_online/model/losses/motion_transvae@criterion  # loss func

personalised/code/configs/logger/none.yaml

@@ -0,0 +1,4 @@
+logger_name: none
+version: ${run_id}
+logger_dir: tb_logs
+project: null

personalised/code/configs/main.yaml

@@ -0,0 +1,26 @@
+
+hydra:
+  run:
+    dir: outputs/${trainer.model_name}/${data.data_name}/${task}/${run_id}
+  job:  # automatically go to the job folder (needed for hydra > 1.2 with new behavior)
+    chdir: true
+
+# Global configurations shared between different modules
+seed: 1234
+logger_level: INFO
+
+task: ???  # online | offline (* mandatory)
+stage: ???  # fit | test (* mandatory)
+data_dir: ???
+# e.g., /lustre/projects/Research_Project-T127204/xk219/projects/datasets/react2026
+resume: false
+run_id: ${generate_id:}  # function defined in prepare.py
+resume_id: ""  # specified if resume
+num_gts: 10  # ground-truths
+
+defaults:
+  - data: ???  # react_2025 (* mandatory)
+  - trainer: ???  # motion_diffusion | motion_transvae (* mandatory)
+  - logger: none
+  - /path@path  # path configuration (loaded from configs/path.yaml)
+  - _self_

personalised/code/configs/model/emotion_autoencoder.yaml

@@ -0,0 +1,17 @@
+type: EmotionVAE
+_target_: framework.modules.emotion_autoencoder.EmotionVAE
+
+in_channels: 25
+out_channels: 25
+feature_dim: 512
+nhead: 8
+dropout: 0.1
+num_encoder_layers: 6
+num_decoder_layers: 6
+mlp_dist: False  # expand mu & logvar
+in_proj_type : linear  # linear | mlp
+out_proj_type : separate  # separate | shared
+pe_type : absolute  # learnable | absolute
+query_type : zero
+max_seq_len: 5000
+global_token_len: 64

personalised/code/configs/model/losses/emotion_autoencoder.yaml

@@ -0,0 +1,6 @@
+_target_: framework.utils.losses.EmotionVAELoss
+
+w_au: 1.0
+w_va: 10.0
+w_em: 1.0
+w_kld: 0.001

personalised/code/configs/model/losses/motion_diffusion.yaml

@@ -0,0 +1,11 @@
+_target_: framework.utils.losses.DiffusionLoss
+
+losses_type: MSELossApt
+losses_multiplier: 1.0
+n_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+temporal_loss_w: 0.0
+prior_loss_weight: 1.0
+eeg_loss_weight: 0.25
+w_au: 1.0  # action unit
+w_va: 5.0  # valence and arousal
+w_em: 5.0  # emotion

personalised/code/configs/model/losses/motion_diffusion_causal.yaml

@@ -0,0 +1,9 @@
+# Loss with the explicit coarse-to-fine cross-entropy term added.
+defaults:
+  - motion_diffusion
+  - _self_
+
+_target_: framework.utils.losses_causal.DiffusionLossCoarse
+w_coarse: 0.5          # weight of the coarse 8-class CE term
+coarse_emo_start: 17
+coarse_classes: 8

personalised/code/configs/model/losses/motion_transvae.yaml

@@ -0,0 +1,8 @@
+_target_: framework.utils.losses.VAELoss
+
+kl_p: 0.00001  # kl div loss
+w_emo: 2.0
+w_exp: 2.0  # facial expression
+w_rot: 4.0  # rotation
+w_tran: 4.0  # translation
+eeg_loss_weight: 0.25

personalised/code/configs/model/losses/perfrdiff_rewrite_weight.yaml

@@ -0,0 +1,8 @@
+type: DiffusionLoss
+args:
+    losses_type: [MSELoss, MSELoss]
+    losses_multipliers: [0, 1]
+    losses_decoded: [False, True]
+    k: ${trainer.${trainer.trainer_mode}.num_preds}
+    temporal_loss_w: 0.0
+    eeg_loss_weight: 1.0

personalised/code/configs/model/motion_diffusion.yaml

@@ -0,0 +1,116 @@
+diff_model:
+    model_name: LatentMLPMatcher
+    _target_: framework.motion_diffusion.diffusion.matchers.LatentMatcher
+
+    eeg_head:
+        enabled: True
+        output_dim: 14
+        hidden_dim: 256
+        dropout: 0.5
+        pooling: mean
+        detach_prediction_emotion: True
+        use_speaker_audio: True
+        use_speaker_emotion: True
+        use_speaker_3dmm: True
+        use_prediction_emotion: True
+
+    diffusion_prior:
+        type:
+            DiffusionPriorNetwork
+        args:
+            emb_preprocessing: normalize
+            freeze_encoder: True
+            audio_dim: 768
+            window_size: ${trainer.${trainer.trainer_mode}.window_size}
+            s_ratio: ${trainer.${trainer.trainer_mode}.s_ratio}
+            token_len: 750
+            _3dmm_dim: 58
+            speaker_emb_dim: 512
+            latent_dim: 512
+            depth: 4
+            num_time_layers: 2
+            num_time_embeds: 1
+            num_time_emb_channels: 64
+            time_last_act: False
+            use_learned_query: True
+            s_audio_cond_drop_prob: 0.2
+            s_latentemb_cond_drop_prob: 1.0
+            s_3dmm_cond_drop_prob: 0.2
+            guidance_scale: 1.0
+            dim_head: 64
+            heads: 8
+            ff_mult: 4
+            norm_in: False
+            norm_out: True
+            attn_dropout: 0.0
+            ff_dropout: 0.0
+            final_proj: True
+            normformer: False
+            rotary_emb: True
+        scheduler:
+            noise_schedule: cosine
+            timestep_spacing: leading
+            num_train_timesteps: 1000
+            num_inference_timesteps: 50
+            predict: start_x
+            var_type: fixed_large
+            rescale_timesteps: False
+            noise_std: 1
+            num_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+
+    diffusion_decoder:
+        type:
+            TransformerDenoiser
+        args:
+            emb_preprocessing: normalize
+            freeze_encoder: True
+            window_size: ${trainer.${trainer.trainer_mode}.window_size}
+            s_ratio: ${trainer.${trainer.trainer_mode}.s_ratio}
+            token_len: 750
+            encode_emotion: False
+            encode_3dmm: False
+            ablation_skip_connection: True
+            nfeats: 25
+            latent_dim: 512
+            ff_size: 1024
+            num_layers: 9  # 7 | 9
+            num_heads: 8  # 4 | 8
+            dropout: 0.1
+            normalize_before: False
+            activation: gelu
+            flip_sin_to_cos: True
+            return_intermediate_dec: False
+            position_embedding: learned
+            arch: trans_dec
+            freq_shift: 0
+            time_encoded_dim: 64
+            s_audio_dim: 768
+            s_audio_scale: 1.0
+            s_emotion_dim: 25
+            l_embed_dim: 512
+            s_embed_dim: 512
+            personal_emb_dim: 512
+            s_3dmm_dim: 58
+            concat: concat_first
+            guidance_scale: 1.0  # 7.5
+            s_audio_enc_drop_prob: 0.2
+            s_latent_embed_drop_prob: 1.0
+            s_3dmm_enc_drop_prob: 0.2
+            s_emotion_enc_drop_prob: 0.2
+            past_l_emotion_drop_prob: 1.0
+            use_past_frames: False
+        scheduler:
+            noise_schedule: cosine
+            timestep_spacing: leading
+            num_train_timesteps: 1000
+            num_inference_timesteps: 50
+            predict: start_x
+            var_type: fixed_large
+            rescale_timesteps: False
+            noise_std: 1
+            num_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+
+defaults:
+    - /model/losses/motion_diffusion@loss
+    - /model/motion_diffusion/audio_embedder@audio_encoder
+    #- /model/latent_embedder@latent_embedder

personalised/code/configs/model/motion_diffusion/audio_embedder.yaml

@@ -0,0 +1,4 @@
+type: AudioEmbedder
+_target_: dataset.modules.audio_embedder.AudioEmbedder
+checkpoint_path: ""
+skip_norm: True

personalised/code/configs/model/motion_diffusion_causal.yaml

@@ -0,0 +1,20 @@
+# Causal / coarse-to-fine variant of the generic motion_diffusion model.
+# Inherits the full baseline model config and only swaps the matcher target
+# and adds the causal-decoder hyperparameters. Non-invasive: originals untouched.
+defaults:
+  - motion_diffusion
+  - _self_
+
+diff_model:
+  _target_: framework.motion_diffusion.diffusion.matchers_causal.CausalLatentMatcher
+  diffusion_decoder:
+    args:
+      # causal lead-lag cross-attention bias
+      use_lag_bias: True
+      lag_max: 60          # max past speaker offset with its own learned bias bucket
+      lag_lookahead: 0     # 0 = strict causal (no future speaker frames)
+      # coarse-to-fine 8-class facial-expression plan
+      use_coarse: True
+      coarse_classes: 8
+      coarse_hidden: 256
+      coarse_emo_start: 17  # emotion 25-d layout: [0:15]=AU, [15:17]=VA, [17:25]=8 expr

personalised/code/configs/model/motion_diffusion_velocity.yaml

@@ -0,0 +1,116 @@
+diff_model:
+    model_name: LatentMLPMatcher
+    _target_: framework.motion_diffusion.diffusion.matchers_velocity.VelocityLatentMatcher
+
+    eeg_head:
+        enabled: True
+        output_dim: 14
+        hidden_dim: 256
+        dropout: 0.5
+        pooling: mean
+        detach_prediction_emotion: True
+        use_speaker_audio: True
+        use_speaker_emotion: True
+        use_speaker_3dmm: True
+        use_prediction_emotion: True
+
+    diffusion_prior:
+        type:
+            DiffusionPriorNetwork
+        args:
+            emb_preprocessing: normalize
+            freeze_encoder: True
+            audio_dim: 768
+            window_size: ${trainer.${trainer.trainer_mode}.window_size}
+            s_ratio: ${trainer.${trainer.trainer_mode}.s_ratio}
+            token_len: 750
+            _3dmm_dim: 58
+            speaker_emb_dim: 512
+            latent_dim: 512
+            depth: 4
+            num_time_layers: 2
+            num_time_embeds: 1
+            num_time_emb_channels: 64
+            time_last_act: False
+            use_learned_query: True
+            s_audio_cond_drop_prob: 0.2
+            s_latentemb_cond_drop_prob: 1.0
+            s_3dmm_cond_drop_prob: 0.2
+            guidance_scale: 1.0
+            dim_head: 64
+            heads: 8
+            ff_mult: 4
+            norm_in: False
+            norm_out: True
+            attn_dropout: 0.0
+            ff_dropout: 0.0
+            final_proj: True
+            normformer: False
+            rotary_emb: True
+        scheduler:
+            noise_schedule: cosine
+            timestep_spacing: leading
+            num_train_timesteps: 1000
+            num_inference_timesteps: 50
+            predict: start_x
+            var_type: fixed_large
+            rescale_timesteps: False
+            noise_std: 1
+            num_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+
+    diffusion_decoder:
+        type:
+            TransformerDenoiser
+        args:
+            emb_preprocessing: normalize
+            freeze_encoder: True
+            window_size: ${trainer.${trainer.trainer_mode}.window_size}
+            s_ratio: ${trainer.${trainer.trainer_mode}.s_ratio}
+            token_len: 750
+            encode_emotion: False
+            encode_3dmm: False
+            ablation_skip_connection: True
+            nfeats: 25
+            latent_dim: 512
+            ff_size: 1024
+            num_layers: 9  # 7 | 9
+            num_heads: 8  # 4 | 8
+            dropout: 0.1
+            normalize_before: False
+            activation: gelu
+            flip_sin_to_cos: True
+            return_intermediate_dec: False
+            position_embedding: learned
+            arch: trans_dec
+            freq_shift: 0
+            time_encoded_dim: 64
+            s_audio_dim: 768
+            s_audio_scale: 1.0
+            s_emotion_dim: 25
+            l_embed_dim: 512
+            s_embed_dim: 512
+            personal_emb_dim: 512
+            s_3dmm_dim: 58
+            concat: concat_first
+            guidance_scale: 1.0  # 7.5
+            s_audio_enc_drop_prob: 0.2
+            s_latent_embed_drop_prob: 1.0
+            s_3dmm_enc_drop_prob: 0.2
+            s_emotion_enc_drop_prob: 0.2
+            past_l_emotion_drop_prob: 1.0
+            use_past_frames: False
+        scheduler:
+            noise_schedule: cosine
+            timestep_spacing: leading
+            num_train_timesteps: 1000
+            num_inference_timesteps: 50
+            predict: start_x
+            var_type: fixed_large
+            rescale_timesteps: False
+            noise_std: 1
+            num_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+
+defaults:
+    - /model/losses/motion_diffusion@loss
+    - /model/motion_diffusion/audio_embedder@audio_encoder
+    #- /model/latent_embedder@latent_embedder

personalised/code/configs/model/motion_transvae.yaml

@@ -0,0 +1,24 @@
+model_name: TransformerVAE
+_target_: framework.motion_transvae.TransformerVAE.TransformerVAE
+
+img_size: 224
+audio_dim: 768
+output_3dmm_dim: 58
+output_emotion_dim: 25
+feature_dim: 128
+seq_len: ${data.train_dataset.clip_length}
+task: ${task}
+window_size: ${trainer.window_size}
+device: cuda
+
+eeg_head:
+    enabled: True
+    output_dim: 14
+    hidden_dim: 256
+    dropout: 0.5
+    pooling: mean
+    detach_prediction_emotion: True
+    use_speaker_audio: True
+    use_speaker_emotion: True
+    use_speaker_3dmm: True
+    use_prediction_emotion: True

personalised/code/configs/model/optim/adam.yaml

@@ -0,0 +1,5 @@
+_target_: torch.optim.Adam
+
+betas: [0.9, 0.98]
+eps: 1e-09
+#lr: 1e-05

personalised/code/configs/model/optim/adamw.yaml

@@ -0,0 +1,7 @@
+_target_: torch.optim.AdamW
+
+#args:
+#lr: 0.0001
+weight_decay: 1e-4
+betas: [0.9, 0.999]  # [0.9, 0.98]
+# eps: 1e-09

personalised/code/configs/model/optim/sgd.yaml

@@ -0,0 +1,5 @@
+_target_: torch.optim.SGD
+
+#lr: 0.001
+weight_decay: 1e-4
+momentum: 0.9

personalised/code/configs/model/scheduler/cosine_annealing.yaml

@@ -0,0 +1,4 @@
+_target_: torch.optim.lr_scheduler.CosineAnnealingLR
+
+eta_min: 0
+last_epoch: -1

personalised/code/configs/path.yaml

@@ -0,0 +1,5 @@
+# path to additional modules, defined in prepare.py
+deps: ${code_path:./deps}
+datasets: ${code_path:./datasets}
+code_dir: ${code_path:}
+working_dir: ${working_path:""}

personalised/code/configs/pers_offline_causal.yaml

@@ -0,0 +1,20 @@
+# Root-level top config for OFFLINE personalized (perfrdiff rewrite-weight) so it
+# composes at the global package (the in-tree personalized_*/ configs get packaged
+# under their subdir name when loaded via --config-name, which breaks overrides).
+# Non-invasive: originals untouched; this only re-hosts the same defaults at root.
+defaults:
+  - /personalized_offline/data/perfrdiff_rewrite_weight@data
+  - /personalized_offline/trainer/perfrdiff_rewrite_weight@trainer
+  - /shared/logger: none
+  - /shared/path@path
+  - _self_
+
+seed: 1234
+logger_level: INFO
+task: offline
+stage: ???
+data_dir: ???
+resume: false
+run_id: ${generate_id:}
+resume_id: ""
+num_gts: 10

personalised/code/configs/pers_online_causal.yaml

@@ -0,0 +1,19 @@
+# Root-level top config for ONLINE personalized (perfrdiff rewrite-weight), mirror of
+# pers_offline_causal.yaml: re-hosts the personalized_online defaults at the global
+# package so --config-name overrides compose correctly. Non-invasive.
+defaults:
+  - /personalized_online/data/perfrdiff_rewrite_weight@data
+  - /personalized_online/trainer/perfrdiff_rewrite_weight@trainer
+  - /shared/logger: none
+  - /shared/path@path
+  - _self_
+
+seed: 1234
+logger_level: INFO
+task: online
+stage: ???
+data_dir: ???
+resume: false
+run_id: ${generate_id:}
+resume_id: ""
+num_gts: 10

personalised/code/configs/personalized_offline/data/perfrdiff_rewrite_weight.yaml

@@ -0,0 +1,135 @@
+data_name: react_2025
+_target_: dataset.perfrdiff_rewrite_weight.PerFRDiffRewriteWeightDataModule
+
+train_dataset:
+    _target_: dataset.perfrdiff_rewrite_weight.PerFRDiffRewriteWeightDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: True
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: train
+    clip_length: 750
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_ref: True
+    load_personality_l: True
+    personal_condition_mode: ${trainer.main_model.args.personal_condition_mode}
+    load_eeg_l: ${trainer.generic.train_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+    normalize_3dmm: standard
+
+validation_dataset:
+    _target_: dataset.perfrdiff_rewrite_weight.PerFRDiffRewriteWeightDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 8
+    root_dir: ${data_dir}
+    split: val
+    clip_length: 750
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_ref: True
+    load_personality_l: True
+    personal_condition_mode: ${trainer.main_model.args.personal_condition_mode}
+    load_eeg_l: ${trainer.generic.train_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+    normalize_3dmm: standard
+
+test_dataset:
+    _target_: dataset.perfrdiff_rewrite_weight.PerFRDiffRewriteWeightDataset
+    batch_size: ${trainer.batch_size}
+    shuffle: False
+    num_workers: 4
+    root_dir: ${data_dir}
+    split: test
+    clip_length: 750
+    target_size: 224
+    crop_size: 224
+    fps: 30
+    audio_feature_type: wav2vec
+    load_video_s: False
+    load_video_l: False
+    load_audio: True
+    load_emotion_s: True
+    load_emotion_l: True
+    load_3dmm_s: True
+    load_3dmm_l: True
+    load_ref: False
+    load_personality_l: True
+    personal_condition_mode: ${trainer.main_model.args.personal_condition_mode}
+    load_eeg_l: ${trainer.generic.eval_eeg}
+    eeg_dir_name: eeg_processed
+    eeg_target_cols:
+        - TP9
+        - AF7
+        - AF8
+        - TP10
+        - Delta_TP9
+        - Theta_TP9
+        - Alpha_TP9
+        - Beta_TP9
+        - Gamma_TP9
+        - Delta_TP10
+        - Theta_TP10
+        - Alpha_TP10
+        - Beta_TP10
+        - Gamma_TP10
+    eeg_channel_scale: 1000.0
+    eeg_use_tar_fallback: True
+    bidirectional: ${trainer.generic.bidirectional}
+    normalize_3dmm: standard
+    num_test_gts: ${num_gts}

personalised/code/configs/personalized_offline/model/losses/motion_diffusion.yaml

@@ -0,0 +1,11 @@
+_target_: framework.utils.losses.DiffusionLoss
+
+losses_type: MSELossApt
+losses_multiplier: 1.0
+n_preds: ${trainer.${trainer.trainer_mode}.num_preds}
+temporal_loss_w: 0.0
+prior_loss_weight: 1.0
+eeg_loss_weight: 0.25
+w_au: 1.0  # action unit
+w_va: 5.0  # valence and arousal
+w_em: 5.0  # emotion