jbegaint / docker-bootstrap.pytorch

bootstrap.pytorch docker image

bootstrap.pytorch docker image

Docker image for bootstrap.pytorch, a high-level extension for deep learning projects with PyTorch (github).

The image is based on the official PyTorch docker image. It provides bootstrap.pytorch and Pillow-SIMD.

Requirements

• docker
• nvidia-docker
• (docker post-install)

Building

git clone https://github.com/jbegaint/docker-bootstrap.pytorch && \
cd docker-bootstrap.pytorch                                    && \
make build

Options:

• the PyTorch docker tag can be set via the PYTORCH_DOCKER_TAG environment variable (see the list of tags on docker hub)

Usage

• use the docker --gpus option to specify which GPUs to use (see the official nvidia doc for some examples)
• the USER_UID and USER_GID environment variables can be set when mounting volumes, to align host and guest filesystem user permissions
• see utils/docker-bootstrap.sh for a full command example

Examples

Interactive python shell

docker run -it --rm bootstrap.pytorch python

PyTorch command

docker run -it                                                        \
           --rm                                                       \
           --gpus '"device=all"'                                      \
           bootstrap.pytorch                                          \
           python -c 'import torch; print(torch.cuda.is_available())'

Calling bootstrap.pytorch

docker run                               \
    -it                                  \
    --rm                                 \
    --shm-size 1G                        \
    -e USER_UID=$(id -u)                 \
    -e USER_GID=$(id -g)                 \
    -v ${HOME}/data:/data/:ro            \
    -v ${PWD}:/home/bootstrap/workspace/ \
    --gpus '"device=all"'                \
    bootstrap.pytorch                    \
    python -m bootstrap.run --help

Full bootstrap.pytorch example

• Clone bootstrap mnist example:

git clone https://github.com/Cadene/mnist.bootstrap.pytorch/ && \
cd mnist.bootstrap.pytorch

• Get docker run wrapper script:

wget https://raw.githubusercontent.com/jbegaint/docker-bootstrap.pytorch/master/utils/docker-bootstrap.sh && \
chmod 755 docker-bootstrap.sh

• Train for one epoch:

./docker-bootstrap.sh python -m bootstrap.run \
        -o mnist/options/adam.yaml            \
        --engine.nb_epochs 1

• List all training options:

./docker-bootstrap.sh python -m bootstrap.run \
        -o mnist/options/adam.yaml            \
        --help

• View results:

./docker-bootstrap.sh python -m bootstrap.compare -d logs/mnist/adam/