Instructions on how to update ADS won't update the Ivadomed package

While testing #620, I encountered an error that was previously fixed on the Ivadomed side. This is because I didn’t update the Ivadomed package in my environment.

Currently, in our docs, it says that we can update ADS by using these commands:

cd axondeepseg
git pull
pip install -e .

However, those instructions won’t update the Ivadomed package. The user would need to do an additional command: pip install ivadomed==<The latest version supported by ADS>.

Question: Will ADS always support the latest Ivadomed version? In this case, we would just need to add pip install ivadomed to the instructions. But if we insist on installing a fixed version, the user would have to check in the environment.yml file which version to install, which might be a bit cumbersome for the user.

[mariehbourget]

Quick update: as I mentioned in last week’s meeting, there is a PR in ivadomed (1000) that I think will help address these issues, and another related PR for documentation (1125).

We discussed it brieftly in ivadomed meeting today and @dyt811 will put together a meeting with ivadomed, SCT and ADS dev to discuss further.

[hermancollin]

Particularly, running pip inside of setup.py is not going to work reliably. The procedural code in setup.py is stripped out when built into a .whl file. It might work, for a while, if you force people to install ADS from source, but it’s not really welcomed by the PyPA people.

@kousu yeah… I read your discussion with PyPA devs and their opinion about it. Very informative btw. I agree that this is not a long-term solution and we might want to avoid merging PR #631 in its current state.

I would be curious to see how other big “pro” python packages deal with these kinds of issues. We can’t be the first group to encounter this torch dependency can of worms.

Does pip install --upgrade -e . solve the original issue?

Hmm it doesn’t work on my side.

I mean in the end, if we can’t find an appropriate solution for this, we could always ask the user to reinstall the whole environment from scratch like previously suggested here (also, from what I understood, this is how sct deals with this). But it still doesn’t fix the fact that we have to account for an ivadomed-compatible torch version