Base Docker image for python projects needing machine learning framework PyTorch package.

This image works as a base on which you can install your smaller python dependencies. PyTorch is rather large package you rather not download every time your requirements.txt changes. If you have many images using PyTorch, using this as the common base image can save a lot of disk space.

Your own requirements.txt should only contain matching versions of packages. To know the packages already installed, run

docker run --rm okoko/python-torch:<tag> pip freeze

The output can be used as content of --constraint file to pip-compile.

The torch builds for +cuNNN are installed from package index https://download.pytorch.org/whl/cuNNN/. This same index needs to be used to install torchvision. It is not included in the image because it is fairly small and has dependencies we rather maintain in the images build on this.

The image build is licensed using BSD-3, but the underlying Debian, Python, PyTorch and any other software have their own licenses. Please take responsibility and use this image according to licenses of the included software.