Upgrade Pytorch
Closed this issue · 7 comments
Pytorch 1.4 has been released. Would it be possibly to adapt the cuda 10 container to it?
As of a4295e0, PyTorch 1.4.0 is now provided in no-cuda, cuda-10.1, and cuda-9.2. The images are still building on Docker Hub so it will probably be an hour or two until they are available to download.
The new images are now live.
Thank you very much. Would it also be possible to switch to the
FROM nvidia/cuda:10.2-base-ubuntu18.04 base image?
Furthermore, could we upgrade python 3.6 to 3.7? 3.8 is not yet supported by Pytorch.
No, I'm hesitant to make changes that big until I have a better system for versioning Docker tags. a) I don't want to break things for people using these as base images, and b) I don't want things to become too unmanageable for myself to keep track of.
I'll probably reevaluate the situation again after Ubuntu 20.04 drops, because I do want to be able to update things more easily moving forward. In the meantime, feel free to just download the Dockerfile you like and tweak it---it should be as easy as just changing the version numbers (I use a bunch of 18.04-based images myself).
All right!
@Zethson I've overhauled the repo and added a few new images for PyTorch 1.5.0 which are based on Ubuntu 18.04. You might be interested in anibali/pytorch:1.5.0-cuda10.2-ubuntu18.04, which also uses Python 3.8. I plan on creating Ubuntu 20.04 variants too as soon as NVIDIA releases appropriate base images.
Thank you very much!