Add interpretability/classifier: standalone set-classifier (train / eval / score)

[al5250]

What

Adds src/ops_model/interpretability/classifier/ — a self-contained toolkit to train, evaluate, and score a permutation-invariant set classifier over single-cell embeddings. It predicts a gene knockout (or a protein-complex / pathway label) from a set ("bag") of cells, and runs independently of the source research repo: the only inputs are an embedding parquet and (for complex classification) a gene→label CSV (shipped), passed as paths.

Files

• train.py — model (MixedChannelClassifier: an inducing-point set transformer that pools all a cell-set's channels into one prediction), the parquet/dump dataset + label-map logic, the training loop, and build_model (checkpoint → model). Hydra CLI.

• eval.py — top-1/top-5 accuracy sweep vs. cells-per-set → curve PNG + JSON + per-class CSV. Hydra CLI.

• score.py — per-cell attribution score (leave-one-out marginal averaged over bag sizes) → ranked CSV. argparse CLI.

• configs/ — 4 train configs + 1 eval config, plus the shipped EBI complex label map ebi_complexes.csv (311 genes → complex/pathway):

  config	task
  train_set_classifier_phase_1K.yaml	phase, 1K gene
  train_set_classifier_phase_ebi.yaml	phase, EBI complex
  train_set_classifier_fluor_1K.yaml	fluor, all markers, 1K gene
  train_set_classifier_fluor_ebi.yaml	fluor, all markers, EBI complex
  eval_set_classifier.yaml	eval (one config for all four)

Usage

pip install -e ".[classifier]"   # hydra-core, omegaconf, pyarrow, tqdm, matplotlib

python -m ops_model.interpretability.classifier.train --config-name train_set_classifier_fluor_1K \
    data.parquet_entries='[{path:/path/to/embeddings.parquet}]' save_path=set_classifier.pt
python -m ops_model.interpretability.classifier.eval  --config-name eval_set_classifier \
    checkpoint_path=set_classifier.pt val_dump_dir=/path/to/dumps/val
python -m ops_model.interpretability.classifier.score --checkpoint set_classifier.pt \
    --dump_dir /path/to/dumps/train /path/to/dumps/val --channel Phase2D --genes KIF23 --out_csv score.csv

The two *_ebi configs point label_map_path at the shipped configs/ebi_complexes.csv. Fluor configs pool all markers per set (channels_per_set: null); set [1, 2, 5, 10, 20, null] to also train on smaller random marker subsets (inline comment in the configs).

Notes

• Mixed-channel only. The classifier pools all of a set's channels into one prediction (mixed_channels_mode: true); the older two-level per-channel path (SetClassifier) was dropped for the release.
• No W&B / cloud coupling. Training prints per-epoch val accuracy to stdout and can persist metrics via metrics_out.
• Extracted ~verbatim from the source set-classifier (only imports, config paths, and dropped W&B/per-channel code differ). Verified equivalent: a full 200-epoch phase-EBI run reproduced the source's val accuracy (best 0.843 vs 0.840), and a 25-epoch run of the trimmed release code reproduced it again (0.749) after the per-channel removal.
• Draft — opening for review; not marking ready yet.

🤖 Generated with Claude Code