PyTorch + Jupyter + GPUs

Creates a docker image for running PyTorch on NVIDIA GPUs with Jupyter notebook support

Structure

• There's a "regular" dockerfile in the jupyter subdirectory. It adds useful python libraries to the pytorch docker image. This is sort of our "base" image.

• There are "fancy" versions in the jupyter_R and jupyter_R_julia subdirectories, which also have R and Julia installed and set up to work with Jupyter. The jupyter_R version is built off of the jupyter image, and jupyter_R_julia is built off of jupyter_R.

• If you just want to run things and not build them yourself, use docker pull rorydm/pytoch_extras:<tag> where <tag> is (say) jupyter_R_julia to pull the prebuilt images from docker hub.

• To build any of the images yourself, from the main directory use bash build_dockerfile.sh <subdir_name>.

• Run any of them from the main directory by using bash run_docker_image.sh <subdir_name> <port>, where <port> is the port you're forwarding out of the docker container, e.g. 9699.

Detailed Installlation Instructions

• On an Ubuntu system (e.g. aws) install current nvidia drivers:

  • sudo add-apt-repository ppa:graphics-drivers/ppa
  • sudo apt-get update && sudo apt-get install nvidia-390

• Install nvida docker (and docker): https://github.com/NVIDIA/nvidia-docker

  • curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
  • curl -s -L https://nvidia.github.io/nvidia-docker/ubuntu16.04/amd64/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list
  • sudo apt-get update
  • sudo apt-get install -y nvidia-docker2
  • sudo pkill -SIGHUP dockerd

• Install pytorch on docker: https://github.com/pytorch/pytorch

  • git clone https://github.com/pytorch/pytorch.git
  • cd pytorch && docker build -t rorydm/pytorch:master .

• Clone this repo next to the pytorch one and run the build script

  • bash build_dockerfile.sh jupyter to first build the minimally extended version
  • bash build_dockerfile.sh jupyter_R to next build the version with R
  • bash build_dockerfile.sh jupyter_R_julia to next build the version with R and Julia

• To run and start a jupyter server (maybe start a screen session first), from this directory:

  • bash run_docker_image.sh <subdir_name> <port>

• To enter the running docker container:

  • docker exec -it <container_name> bash

Note: the terminal inside docker sometimes acts funny. using the following seems to help:

• docker exec -t <container_name> /bin/bash -c "export COLUMNS=tput cols; export LINES=tput lines; exec bash"