Decide on using UV/python venv/conda/etc., package manager

[c-salomonsen]

Possible choices that could work:

• UV
• venv (requires no other dependencies)
• conda (mamba)
• Others?

[Seilmast]

We should have a docker image we can pull from. The cluster can pull from it, making the individual use of package installer not something we need to think about.
If I get the packages we need, I can try setting up the image

[c-salomonsen]

Good idea.

Check out this PR #5, I've added an initial environment.yml, but its only for Sphinx related stuff, which we probably wont need in the dockerfile, but then we would need to balance possibly 2 different requirement files.

[Johanmkr]

Any progress with the dockerfile/dockerimage? I am testing with miniconda but had to add several packages to the environment.yml file

[c-salomonsen]

@Johanmkr I can have a go at the dockerfile/image later today, maybe setting up some github action for building the image(?).

What dependencies do you have? Maybe a branch?

[Seilmast]

I was supposed to make the Dockerfile on friday, but I got sick Tuesday evening. I'm still ill, so if someone want to push a docker image and provide the tag feel free to do so.
If not, I'll see if I can do it later tonight or tomorrow.

[Johanmkr]

@salomaestro I added torch and torchvision, but I see that you did the same #27 so there should not be a problem. I am still just working on my own branch.

[Seilmast]

I've made a docker image under the tag
seilmast/colabexam:latest
Can anyone confirm they can pull the image?

[c-salomonsen]

@Seilmast Yes, I was able to pull the image locally 👍🏻

[Johanmkr]

@Seilmast I am able to pull, but is it supposed to be >5GB?

[c-salomonsen]

@Johanmkr I think if pytorch and torchvision is installed, that image size is to be expected. I wasn't able to verify myself what is installed due to being on Macos, but you should be able to verify it locally with docker run seilmast/colabexam:latest python -c "import torch; print(torch.tensor(1))".

[Seilmast]

@Johanmkr As @salomaestro said, it's mostly torch. I pull the latest torch image from docker, so you only have to pull the image when we update it. It shouldn't affect the speed of your code that much.
Torch, as the whole package, is maybe around 3-4GB I'd guess. Might be more.