Manvith/Endoscopy-Eso

NOVA-GI Unified — Inference App

A dependency-light inference UI for the trained NOVA-GI Unified models. It reproduces the exact preprocessing and pipeline from NOVA_GI_Unified_updated.ipynb (final inference cell).

What it does

• —Loads all saved dataset models (one head per dataset — Gastroendonet, Gastrovision, hyper-kvasir, kvasir-v1, kvasir-v2, WCEBleedGen). For each it picks best_ema.pt or best_model.pt per the run's chosen_model, plus its class_map.json + temperature.json.
• —For every uploaded image it applies that dataset's preprocessing (specular-inpaint → gray-world → CLAHE-LAB → resize 240px INTER_AREA → ToTensor → ImageNet normalize), runs the full architecture, and returns softmax(logits / T).
• —All heads score the image; the most confident findings are ranked (class / sub-class / region / severity / calibrated confidence).
• —One-click Save interactive HTML exports a self-contained report.

Run locally

.venv\Scripts\python.exe inference_app\app.py

Then open <http://127.0.0.1:7860>. Uses the local ../NOVA_GI_Unified_outputs/<dataset>/seed42_var_full/ weights. Runs on CPU.

Deploy (Hugging Face Space)

Weights are not committed to this repo. Set the NOVA_GI_MODEL_REPO environment variable (Space → Settings → Variables) to your HF model repo id (e.g. your-username/nova-gi-unified). On boot the app downloads the six checkpoints + JSONs from that repo and caches them. See DEPLOY.md for the full step-by-step, and upload_weights.py for the one-time weight upload.

Files

• —nova_gi_engine.py — model definition, preprocessing, registry, predict_image().
• —app.py — stdlib web server + browser UI + HTML export.
• —Dockerfile, requirements.txt — container build for the Space.
• —upload_weights.py — pushes the needed checkpoints to an HF model repo.

Notes

• —Research/educational use only — not a medical device.