Dockerfile.base

# Base image for debugging Python environment
# Build: docker build -f Dockerfile.base -t cvat.pth.mmseg.coralscopsegformer.base:latest .
# Debug: docker run -it --gpus all cvat.pth.mmseg.coralscopsegformer.base:latest python -c "from main import handler; print('OK')"

FROM pytorch/pytorch:2.7.0-cuda12.8-cudnn9-runtime

# Environment Variables
ENV NVIDIA_VISIBLE_DEVICES=all \
    NVIDIA_DRIVER_CAPABILITIES=compute,utility \
    DEBIAN_FRONTEND=noninteractive

# System Dependencies
RUN apt-get update && apt-get install -y \
    git \
    libgl1-mesa-glx \
    libglib2.0-0 \
    --no-install-recommends && \
    rm -rf /var/lib/apt/lists/*

# Work Directory
WORKDIR /opt/nuclio

# Python Package Installation
# Note: Using custom index for CUDA 12.8 compatible mmcv
RUN pip install --extra-index-url https://miropsota.github.io/torch_packages_builder \
    mmcv==2.2.0+pt2.7.0cu128 mmengine mmsegmentation>=1.0.0 torch==2.7.0

# Install DINOv2 without dependencies to bypass version conflicts
RUN pip install --no-cache-dir git+https://github.com/facebookresearch/dinov2.git --no-deps

RUN pip install --no-cache-dir \
    fvcore==0.1.5.post20221221 \
    iopath==0.1.10 \
    omegaconf==2.3.0 \
    submitit==1.5.1 \
    torchmetrics==1.4.0 \
    xformers==0.0.30 ftfy regex scikit-learn matplotlib pycocotools

# Workaround for MMCV version compatibility
ENV TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1
RUN sed -i "s/MMCV_MAX = '2.2.0'/MMCV_MAX = '2.3.0'/" /opt/conda/lib/python3.11/site-packages/mmseg/__init__.py

# Copy application code
COPY . .

# Download pre-trained models
ADD https://storage.googleapis.com/public_model_repo/dinov2_vitb14_coralsegv4_ms_config_segformer/dinov2_vitb14_coralsegv4_ms_config_segformer.py ./best_segformer_config.py
ADD https://storage.googleapis.com/public_model_repo/dinov2_vitb14_coralsegv4_ms_config_segformer/best_mIoU_epoch_140.pth ./best_segformer.pth
ADD https://storage.googleapis.com/public_model_repo/vit_b_coralscop.pth ./