Any ideas regarding reducing Image size? about docker-pytorch HOT 1 CLOSED

Multi-stage builds are useful for situations where you need to build software that is included in the image, since you can setup a development environment in one image, build the software there, then copy the built artifacts across to the actual final image, all in one Dockerfile. However, if you look at e.g. https://github.com/anibali/docker-pytorch/blob/master/dockerfiles/1.7.0-cuda11.0-ubuntu20.04/Dockerfile you'll notice that it just uses binaries and doesn't need to build anything. So I don't see a way of using multi-stage builds to reduce the size of the base docker-pytorch image. Perhaps it could be useful if any of the additional software you're adding requires building.

Unfortunately, I can't see much fat to cut at all from my base images---PyTorch and its dependencies are just really big. I suppose if you don't need torchvision you could cut that from the image yourself.

from docker-pytorch.