Dockerfile

# Base image with PyTorch and CUDA for GPU support
FROM pytorch/pytorch:1.10.0-cuda11.3-cudnn8-runtime

# Install system dependencies and Hugging Face CLI
RUN apt-get update && apt-get install -y \
    git \
    wget \
    && rm -rf /var/lib/apt/lists/*

# Install Python packages including Hugging Face Transformers, TorchScript, and Flask
RUN pip install --no-cache-dir \
    torch \
    torchvision \
    transformers \
    requests \
    Flask \
    Pillow

# Set Hugging Face cache to a guaranteed writable directory
ENV TRANSFORMERS_CACHE=/tmp/cache
RUN mkdir -p /tmp/cache

# Create directories for the models
RUN mkdir -p /models/sapiens_pose /models/motionbert

# Python script to download models
RUN echo "from transformers import AutoModel, AutoTokenizer\n\
AutoModel.from_pretrained('facebook/sapiens-pose-1b-torchscript').save_pretrained('/models/sapiens_pose')\n\
AutoModel.from_pretrained('walterzhu/MotionBERT').save_pretrained('/models/motionbert')\n\
AutoTokenizer.from_pretrained('walterzhu/MotionBERT').save_pretrained('/models/motionbert')" > download_models.py

# Run the script to download models
RUN python download_models.py

# Copy the inference script (app.py) into the container
COPY app.py /app/app.py

# Expose the default port for Flask
EXPOSE 7860

# Run the Flask app
CMD ["python", "/app/app.py"]