Fix: Clone SAM2 from Meta repo and download weights using official scripts

Dockerfile

@@ -1,4 +1,4 @@
-FROM python:3.11-slim
+FROM pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime
 
 WORKDIR /app
 
@@ -6,42 +6,62 @@ WORKDIR /app
 RUN apt-get update && apt-get install -y \
     git \
     wget \
+    curl \
     libsm6 \
     libxext6 \
     libxrender-dev \
     libgl1-mesa-glx \
     && rm -rf /var/lib/apt/lists/*
 
-# Copy requirements first for better caching
+# Install Python dependencies first
 COPY requirements.txt .
-
-# Install Python dependencies
 RUN pip install --no-cache-dir -r requirements.txt
 
-# Copy model checkpoints and application files
-COPY segment-anything-2/ ./segment-anything-2/
+# Clone Segment Anything 2 from Meta
+RUN git clone https://github.com/facebookresearch/sam2.git segment-anything-2 && \
+    cd segment-anything-2 && \
+    pip install -e . && \
+    cd ..
+
+# Download SAM2 model weights
+RUN cd segment-anything-2/checkpoints && \
+    bash download_ckpts.sh && \
+    cd ../..
+
+# Set Hugging Face token for VREyeSAM weights (injected by HF Spaces)
+ENV HF_TOKEN=""
+
+# Copy application files
 COPY app.py model_server.py ./
 
 # Create .streamlit directory and config
 RUN mkdir -p .streamlit
 COPY .streamlit/config.toml .streamlit/
 
-# Set environment variable for Hugging Face authentication
-ENV HF_TOKEN=""
-
-# Verify model files exist
-RUN python -c "
+# Download VREyeSAM fine-tuned weights from Hugging Face
+# This assumes your weights are available on HF Model Hub
+RUN if [ ! -z "$HF_TOKEN" ]; then \
+    python -c "
+from huggingface_hub import hf_hub_download
 import os
-paths = [
-    'segment-anything-2/checkpoints/sam2_hiera_small.pt',
-    'segment-anything-2/checkpoints/VREyeSAM_uncertainity_best.torch'
-]
-for p in paths:
-    if not os.path.exists(p):
-        print(f'Warning: {p} not found - will be downloaded at runtime')
-    else:
-        print(f'Found: {p}')
-"
+
+hf_token = os.getenv('HF_TOKEN')
+if hf_token:
+    # Download VREyeSAM weights from your HF repo
+    # Replace 'your-username/vreyesam' with your actual repo
+    try:
+        checkpoint = hf_hub_download(
+            repo_id='devnagaich/VREyeSAM',
+            filename='VREyeSAM_uncertainity_best.torch',
+            token=hf_token,
+            cache_dir='segment-anything-2/checkpoints'
+        )
+        print(f'Downloaded VREyeSAM weights to: {checkpoint}')
+    except Exception as e:
+        print(f'Warning: Could not download VREyeSAM weights: {e}')
+        print('App will attempt to load from cache at runtime')
+" || true; \
+    fi
 
 # Expose Streamlit port
 EXPOSE 7860

requirements.txt

@@ -2,8 +2,8 @@
 streamlit>=1.28.0
 
 # Core ML and Deep Learning
-torch>=2.0.0,<2.5.0
-torchvision>=0.15.0,<0.20.0
+torch>=2.5.1,<2.6.0
+torchvision>=0.20.1,<0.21.0
 numpy>=1.22.0,<2.0.0
 
 # Computer Vision
@@ -17,5 +17,8 @@ scikit-learn>=1.0.0
 # Visualization
 matplotlib>=3.5.0
 
+# Hugging Face integration
+huggingface-hub>=0.19.0
+
 # Utility packages
 tqdm>=4.60.0