Skip to content

README.md

Latest commit

 

History

History
53 lines (38 loc) · 1.58 KB

README.md

File metadata and controls

53 lines (38 loc) · 1.58 KB

SCARCE-CXR

Blog License: MIT

SSL pretraining on 112k unlabeled NIH chest X-rays, evaluated on 10 rare PadChest diseases at 1 to 50 shots. Four methods: MoCo v2, BarlowTwins, SparK, DINO.

Setup

Requires Python 3.14 and uv.

git clone https://github.com/ryanzhou147/SCARCE-CXR && cd SCARCE-CXR
uv sync
uv run python -m data.download --dataset nih
uv run python -m data.download --dataset padchest

Pretraining

uv run python main.py pretrain-moco   --config configs/moco.yaml    # ResNet50, ~63h
uv run python main.py pretrain-barlow --config configs/barlow.yaml  # ResNet18, ~15h
uv run python main.py pretrain-spark  --config configs/spark.yaml   # ResNet50, ~20h

Evaluation

uv run python -m finetune.probe    --checkpoint outputs_cloud/moco-v2/best.pt
uv run python -m finetune.finetune --checkpoint outputs_cloud/moco-v2/best.pt
uv run python -m finetune.count_labels
uv run python -m data.eval.collapse_monitor --outputs-dir outputs/moco-v2
bash run_eval.sh

Structure

ssl_methods/    MoCo, BarlowTwins, SparK, DINO training loops
finetune/       probe.py, finetune.py, _data.py, _plots.py
data/           RAM-cached loader, backbone loader, download, collapse monitor
configs/        per-method YAML hyperparameters
results/        eval summaries