Spaces:
Sleeping
Sleeping
Commit History
-
fix: pin numpy to 1.26.4 — numpy 2.x breaks torchvision 0.17.2 ToTensor
1763697 -
fix: catch all inference exceptions and OSError in SSRF check to surface real 500 cause
cfd4ead -
fix: read CORS allowed origins from ALLOWED_ORIGINS env secret
3fc3b69 -
feat: add Dockerfile and HF Spaces config for deployment
- CPU-only PyTorch for smaller image (~1.5GB vs 6GB) - Downloads model weights from HuggingFace at build time - Separate requirements-deploy.txt (excludes torch/torchvision, installed separately) - HF Spaces YAML frontmatter in README.mdf192d59 -
feat: integrate real SensiNet mammography model
- Add SensiNet dual-stream architecture (Xception + EfficientNet-B3 with CBAM) - Replace mock inference with real Bayesian MC-Dropout prediction (10 passes) - Add /analyze endpoint with SSRF protection (allowlist + private IP blocking) - Add malignancy_probability to response schema - Add training and data preparation utility scripts - Model weights (131MB .pth) gitignored — must be downloaded separately596aaa1 -
feat: bootstrap mammogram inference API starter
1d745c0