billchenxi/surgical-workflow-models

Surgical workflow models

A collection of checkpoints from the Chen lab covering surgical-workflow analysis tasks. This repository is intended to host weights for several papers and projects; new project subfolders will be added over time.

Project: rsd-mb140-cholec80-2026 (current)

Trained model checkpoints for the NeurIPS 2026 submission "When Workflow Conditioning Helps... A Variability-Scaling Study on MultiBypass140 and Cholec80."

12 PyTorch state-dicts produced by the training scripts in the bariatric_rsd GitHub repo:

Cholec80 (used for the §6.5 ensemble + isotonic 3.56 result)

• cholec80/run018_seed42.pth
• cholec80/run018_seed123.pth
• cholec80/run018_seed777.pth

MultiBypass140 fold 0 strict prefix-only (Run 033, §6.1 headline)

• mb140/run033_no_token_seed{42,123,777}.pth
• mb140/run033_oracle_seed{42,123,777}.pth
• mb140/run033_decoupled_seed{42,123,777}.pth

Each checkpoint is the best-val-MAE state-dict from a 15-epoch training run (cosine LR schedule, AdamW, batch 64, sequence_len=8, frame_stride=5).

How to use

See the reproducibility package in the GitHub repo: https://github.com/billchenxi/bariatric_rsd/tree/master/reproducibility

Quick verification (requires Cholec80 frames extracted; see DATA_PREP.md):

git clone https://github.com/billchenxi/bariatric_rsd.git
cd bariatric_rsd/reproducibility
pip install -r requirements.txt
bash scripts/download_weights.sh --only cholec80
CHOLEC80_ROOT=/path/to/cholec80 bash scripts/verify_cholec80_3.56.sh
# Expected: 3.563 ± 0.01 min

Model architecture

• Visual encoder: ViT-B/16 (timm vit_base_patch16_224), ImageNet pretrained, layers 0-5 frozen.
• Temporal head: Hierarchical Temporal Attention (HTA) adapted from Surgformer (Yang et al., 2024).
• Heads: RSD regression, phase classification (auxiliary), deviation classification (auxiliary).
• Workflow conditioning: prepended workflow-cluster token (oracle, decoupled-oracle, or none) at the temporal-head input.

Total parameters: ~93 M (ViT-B + temporal heads).

Caveats and intended use

• These checkpoints are released for academic reproducibility of the NeurIPS 2026 submission. Not intended for clinical use.
• The 3.56 min Cholec80 result depends on a 3-seed ensemble + horizontal- flip TTA + per-video isotonic post-processing. Isotonic post-processing alone accounts for ~83% of the gain — see the manuscript §6.5 for the full caveats.
• Cholec80 results are on a 30-video public phase-labeled subset, not the canonical 40-video test split.
• MultiBypass140 numbers are validation-set (training-time best-checkpoint selection); 5-fold extension is in flight.

License

MIT — same as the GitHub repo. Note that the underlying datasets (Cholec80, MultiBypass140) have their own licenses and are not redistributed here.

Citation

@inproceedings{bariatric_rsd_2026,
  title  = {When Workflow Conditioning Helps: A Variability-Scaling Study on MultiBypass140 and Cholec80},
  author = {Chen, Bill and others},
  booktitle = {Submitted to NeurIPS 2026 (Evaluations \& Datasets track)},
  year   = {2026},
  url    = {https://github.com/billchenxi/bariatric_rsd}
}