minimrbanana / docker-nvidia-tutorial

Docker Nvidia Tutorial

Running a CUDA Capable Docker Container

• Official NVIDIA PyTorch docker container: https://ngc.nvidia.com/catalog/containers/nvidia:pytorch
  • Latest image (as of Aug. 2020): nvcr.io/nvidia/pytorch:20.07-py3
• Show available GPUs:

  docker run --gpus all -it --rm nvcr.io/nvidia/pytorch:20.07-py3 nvidia-smi
      

• See if PyTorch can detect CUDA:

  docker run --gpus all --rm nvcr.io/nvidia/pytorch:20.07-py3 \
      python -c "import torch; print(torch.cuda.is_available())"
      

Creating a Custom Docker Image

• A docker image represents your whole experiment environment, that is
  • Linux environment
  • Python environment
• The Dockerfile gives instructions on how to create the environment

  # Select the base image
  FROM nvcr.io/nvidia/pytorch:20.07-py3
  
  # Select the working directory
  WORKDIR /app
  
  # Install SPFlow from the master branch
  RUN git clone https://github.com/SPFlow/SPFlow && \
      cd SPFlow/src && \
      bash create_pip_dist.sh && \
      pip install dist/spflow-0.0.40-py3-none-any.whl
  
  # Install Python requirements
  COPY ./requirements.txt ./requirements.txt
  RUN pip install -r requirements.txt
      

• To build a docker image (we now name it docker-tutorial), run:

  docker build -t docker-tutorial .
      

• To test the docker image, run:

  docker run --gpus all -it --rm docker-tutorial python
      

Running a Project in the Container

• Project source code is not supposed to be built into the docker image
• Rather, attach all source code as volume when starting a new container:

  Use the --volume SOURCE:DESTINATION to mount a file/directory from the host machine into the docker container:

  docker run --gpus all --rm -it \
      --volume "$(pwd)"/src:/app/src \
      docker-tutorial \
      python /app/src/run.py