ego_gen_train_code

Training code for 2-view egocentric video generation — a fine-tune of Cosmos Predict 2.5 (2B multiview DiT) that jointly generates two synchronized ego views of a 2-person interaction, conditioned on warped scene RGB, per-person skeleton pose, shared reference frames, Plücker camera rays, and (optionally) composite scene+human depth.

This repo is self-contained enough to resume training on a fresh server: it holds the full training code under cosmos-predict2.5/ and the actor-observer warm-start checkpoint under checkpoints/actorobs_shared_iter5500/. Datasets live in companion HF repos (see below).

---

⚡ Quickstart — run CoMind, warm-started from the actor-observer checkpoint, on a new server

This reproduces the comind_actoractor_personpose_shared run: CoMind 2-person actor-actor, person-ID pose, shared refs + Plücker, warm-started from the Nymeria actor-observer checkpoint (only the new conditioning is learned; no depth).

# ---- 0. get this repo (code + warm-start checkpoint) ----
pip install -U "huggingface_hub[cli]"
huggingface-cli download dhyun22/ego_gen_train_code --repo-type dataset \
  --local-dir ~/ego_gen_train_code
cd ~/ego_gen_train_code/cosmos-predict2.5

# ---- 1. environment (Python 3.10 + CUDA 12.8, FSDP2) ----
uv sync --frozen                      # reproduces the exact env from uv.lock into ./.venv
#   (or recreate the origin conda env `ego_dh`; then point PY= in sh/train_nymeria_longer.sh at it)

# ---- 2. CoMind dataset -> the EXACT path the config expects ----
huggingface-cli download dhyun22/ego_gen_comind --repo-type dataset \
  --local-dir /data4/comind_dataset

# ---- 3. actor-observer warm-start checkpoint -> the EXACT path the config expects ----
#     (it ships inside THIS repo under checkpoints/; copy it to the load_path the experiment hardcodes)
mkdir -p /data/model_output/warmstart
cp -r ~/ego_gen_train_code/checkpoints/actorobs_shared_iter5500 \
      /data/model_output/warmstart/actorobs_shared_iter5500

# ---- 4. base backbone + text tokenizer (public NVIDIA repos) ----
#   base 2B DiT: auto-downloaded on first run from `nvidia/Cosmos-Predict2.5-2B`
#   text tokenizer (16GB): download once and point COSMOS_QWEN_TOKENIZER_DIR at it
huggingface-cli download nvidia/Cosmos-Reason1-7B --local-dir /data/cosmos_reason1_7b

# ---- 5. secrets (not in this repo) ----
cat > ~/.ego_dh_secrets.sh <<'EOF'
export HF_TOKEN=hf_xxx            # HF pulls
export WANDB_API_KEY=xxx          # W&B logging (or unset to disable)
EOF
#   the launch script sources /home/ubuntu/.ego_dh_secrets.sh — put it there, or edit the path in sh/train_nymeria_longer.sh

# ---- 6. launch (4-GPU FSDP; DCP checkpoint reshards, so any NPROC works) ----
EXP=comind_actoractor_personpose_shared NPROC=4 PORT=29500 bash sh/train_nymeria_longer.sh

Why this "just works": the experiment comind_actoractor_personpose_shared hardcodes checkpoint.load_path = /data/model_output/warmstart/actorobs_shared_iter5500 and the CoMind data root /data4/comind_dataset. Steps 2–3 put the files at exactly those paths, so no overrides are needed. If you must use different paths, override on the CLI, e.g. append checkpoint.load_path=/my/ckpt to the launch command.

Expected sanity signals: iter-1 Loss ≈ 0.05–0.07 (low, because it warm-starts from the trained actor-observer). Checkpoints + validation videos land in /data/model_output/train/comind_actoractor_personpose_shared/; the run auto-resumes if that dir already has checkpoints/latest_checkpoint.txt. W&B run name == the experiment name.

Want the CoMind + synthetic mix instead? Use EXP=comind_synth_actoractor_personpose_shared and also download the synthetic set to /data3/synthetic_processed_multi. Everything else is identical.

---

Repository layout

cosmos-predict2.5/
  sh/train_nymeria_longer.sh              # launch script (torchrun + hydra)
  uv.lock                                 # pinned dependencies
  cosmos_predict2/_src/predict2_multiview/
    configs/vid2vid/experiment/nymeria_pose_2actor.py   # ALL experiments registered here
    configs/vid2vid/defaults/conditioner.py             # conditioning (pose/warp/refs/plucker/depth)
    networks/multiview_pose_dit.py                       # DiT + pose/plucker/depth embedders
    models/multiview_pose_model_rectified_flow.py        # rectified-flow model, VAE encode, trainable set
    datasets/nymeria_pairs.py                            # nymeria(longer/single)+synthetic loaders
    datasets/comind_pairs.py                             # CoMind actor-actor loader
    callbacks/nymeria_validation_viz.py                  # validation grid (warped|pose|GT|gen|depth)
checkpoints/
  actorobs_shared_iter5500/model/*.distcp + .metadata    # actor-observer warm-start (DCP, model-only, 3.9GB)

---

External artifacts (fetched in the Quickstart, summarized)

Artifact	Path the code expects	Source
CoMind dataset	/data4/comind_dataset	HF dhyun22/ego_gen_comind (public)
Nymeria longer (actor-actor)	/data2/nymeria_processed_longer	HF dhyun22/ego_gen_nymeria_longer (if used)
Synthetic multi	/data3/synthetic_processed_multi	companion upload (only for the +synth runs)
actor-observer warm-start	/data/model_output/warmstart/actorobs_shared_iter5500	in THIS repo checkpoints/
base 2B DiT	auto (HF cache)	nvidia/Cosmos-Predict2.5-2B (auto/multiview/..._ema_bf16.pt)
text tokenizer (16GB)	COSMOS_QWEN_TOKENIZER_DIR=/data/cosmos_reason1_7b	nvidia/Cosmos-Reason1-7B

Each CoMind clip provides the training-required arrays (<clip>.npz = RGB + warped_cond + pose + pose_person; <clip>.refs.npz / shared refs; <clip>.campose.npz) plus a manifest/. Intermediate artifacts (raw depth pools, meshes/masks) are not needed at train time.

---

Registered experiments (in nymeria_pose_2actor.py)

EXP	Data	Warm-start	Notes
comind_actoractor_personpose_shared	CoMind	actor-observer ckpt	the Quickstart run; person-ID pose, shared refs; no depth
comind_synth_actoractor_personpose_shared	CoMind + synthetic	actor-observer ckpt	same + synthetic mix
nymeria_synth_actoractor_depth_shared	Nymeria longer + synthetic	actor-observer ckpt	adds composite-depth on the VAE channel axis
nymeria_actorobs_depth_shared	Nymeria single (actor-observer)	fresh from Cosmos base 2B	stage-1 depth run trained from scratch

Launch any with EXP=<name> NPROC=<n> PORT=<p> bash sh/train_nymeria_longer.sh; extra hydra overrides pass through (e.g. trainer.max_iter=5). Outputs → $IMAGINAIRE_OUTPUT_ROOT/train/<name>/ (default /data/model_output).

---

Conditioning summary (what the model consumes)

• warped_cond — scene RGB warped from a combined own+partner past-frame pool (max-coverage source per target).
• pose / pose_person — per-person skeleton render; pose_person uses a view-invariant identity palette (same person → same color across both views).
• refs_shared — one greedy set-cover reference set (V×R clean frames) identical in both views, injected as extra self-attention tokens with a custom RoPE and a zero-init ref_gate.
• campose (Plücker) — per-pixel ray dir+moment in a per-pair canonical frame (view0/frame0), additive zero-init embedder.
• depth (optional) — mesh_cond.npz['depth_comp']: warped scene depth (person-masked) + human mesh depth hole-fill, log-normalized turbo RGB; encoded through the frozen VAE and added via a zero-init depth_embedder.

All new conditioning modules are zero-init / missing-key, so any experiment warm-starts cleanly — from the actor-observer checkpoint (Quickstart) or from the Cosmos base 2B backbone.