Skip to content

Latest commit

 

History

History
66 lines (48 loc) · 2.68 KB

README.md

File metadata and controls

66 lines (48 loc) · 2.68 KB

Using Pyro Docker

Some utilities for building docker images and running Pyro inside a Docker container are included in the docker directory. This includes a Dockerfile to build PyTorch and Pyro, with some common recipes included in the Makefile.

Dependencies for building the docker images:

  • docker (>= version 17.05)
  • nvidia-docker Refer to the readme for installation.

Building Images

The Makefile can be used to build CPU and CUDA images for Pyro and PyTorch. Some common options are as follows:

  1. Source: Uses the latest released package (conda package for PyTorch and PyPi wheel for Pyro) by default. However, both Pyro and PyTorch can be built from source from the master branch or any other arbitrary branch specified by pytorch_branch and pyro_branch.
  2. CPU / CUDA: make build or make build-gpu can be used to specify whether the CPU or the CUDA image is to be built. For building the CUDA image, nvidia-docker is required.
  3. Python Version: Python version can be specified via the argument python_version.

For example, the make command to build an image that uses Pyro's dev branch over PyTorch's master branch, using python 3.6 to run on a GPU, is as follows:

make build-gpu pyro_branch=dev pytorch_branch=master python_version=3.6

This will build an image named pyro-gpu-dev-3.6. To spin up a docker container from this image, and run jupyter notebook on this, use the following make command:

make notebook-gpu img=pyro-gpu-dev-3.6

For help on the make commands available, run make help.

NOTE (Mac Users): Please increase the memory available to the Docker application via Preferences --> Advanced from 2GB (default) to at least 4GB prior to building the docker image (specially for building PyTorch from source).

Running the Docker container

Once the image is built, the docker container can be started via make run, or make run-gpu. By default this starts a bash shell. One could start an ipython shell instead by running make run cmd=ipython. The image to be used can be specified via the argument img.

To run a jupyter notebook use make notebook, or make notebook-gpu. This will start a jupyter notebook server which can be accessed from the browser using the link mentioned in the terminal.

Note that there is a shared volume between the container and the host system, with the location $DOCKER_WORK_DIR on the container, and $HOST_WORK_DIR on the local system. These variables can be configured in the Makefile.