Spaces:
Running
Running
Commit History
21 days ago
-
feat: add BiomedCLIP attribute classifier as stage 2 for Cancer predictions Running
/predict now runs the ConvNeXt binary classifier first and, only when it predicts Cancer, runs the BiomedCLIP ViT-B/16 dual-scale attribute model (7 radiological findings with per-label OOF thresholds, no TTA). The attribute model was trained on cancer cases only, hence the gating. - attributes.py rebuilds BiomedCLIP's vision tower directly in timm, so no open_clip/transformers dependency or runtime hub download; verified bit-identical logits against the notebook's open_clip construction - inference runs in a worker thread so slow requests don't block /health - attributes returned in JSON, or in the X-Attributes header with gradcam - a stage 2 failure is reported in attributes_error without losing the stage 1 result Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VDHZZbNXTVfMbJQPTS5GRmdce349c
22 days ago
-
feat: switch model from DenseNet121-CBAM to ConvNeXt-Tiny + CBAM + MSAM
Replace the DenseNet121-CBAM classifier with the 5-fold CV ConvNeXt-Tiny + CBAM + MSAM model (fold 3 checkpoint: test AUC 0.938, val F1 0.901). Threshold 0.62, Grad-CAM++ now targets the MSAM layer, and the old DenseNet weights are removed. API response format is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VDHZZbNXTVfMbJQPTS5GRmbfd7148
-
chanegs in the model
f439bb2
-
model updated
f4ea9a8
-
fixes to file name
2a51f8d