Add CheXbert: clone repo in Docker, download weights from HF Hub at startup

Dockerfile

@@ -16,11 +16,20 @@ RUN pip install --no-cache-dir -r requirements.txt
 
 RUN python -c "import nltk; nltk.download('punkt'); nltk.download('punkt_tab')"
 
+# Clone Stanford CheXbert (public repo — only source code, no weights)
+RUN git clone --depth 1 https://github.com/stanfordmlgroup/CheXbert.git /app/CheXbert
+
+# Install CheXbert's own dependencies
+RUN pip install --no-cache-dir scikit-learn tqdm || true
+
 COPY inference/ ./inference/
 COPY main.py .
 
 RUN mkdir -p /tmp/results
 
+ENV CHEXBERT_DIR=/app/CheXbert
+ENV CHEXBERT_CKPT=/app/CheXbert/src/chexbert.pth
+
 EXPOSE 7860
 
 CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "7860"]

main.py

@@ -38,6 +38,23 @@ app.add_middleware(
 
 @app.on_event("startup")
 async def load_model():
+    import os
+    from huggingface_hub import hf_hub_download
+
+    # Download chexbert.pth if not already present
+    chexbert_ckpt = os.environ.get("CHEXBERT_CKPT", "/app/CheXbert/src/chexbert.pth")
+    if not os.path.exists(chexbert_ckpt):
+        print("📥 Downloading chexbert.pth from HuggingFace Hub...")
+        os.makedirs(os.path.dirname(chexbert_ckpt), exist_ok=True)
+        hf_hub_download(
+            repo_id="alyrraza/radguard-v11",
+            filename="chexbert.pth",
+            local_dir=os.path.dirname(chexbert_ckpt),
+        )
+        print("✅ chexbert.pth downloaded")
+    else:
+        print("✅ chexbert.pth already present")
+
     from inference.model import get_model, get_tokenizer
     get_model()
     get_tokenizer()