Specify torch version, which fixed #199

[timephy]

Should fix an error which occurs sporadically when training.

[timephy]

Hey @asogaard,
Here is the torch version fix which resolves #199.
I couldn't specify it in envs/gnn_py38.yml as you seem to have moved it to setup.py (78ccf71).
Not sure why torch is not listed in INSTALL_REQUIRES but in the try-expect, but that is where I changed the version. Hope this works!

[asogaard]

Hi Tim,
Thanks for confirming that this fixes the issue from #199! And to your two comments above:

• I have moved pytorch from the conda environments file to have all python packages installed via pip rather than having some installed through conda and some through pip. The latter could lead to problems when trying to change torch versions after installation.
• The reason for installing pytorch in the try-except block is that pytorch and pytorc-geometric do not play nicely together. This means that pytorch needs to be installed before installing pytorch-geometric. It's ugly, but it's a known design quirk of pytorch-geometric (see: Cannot install torch-geometric via requirement.txt pyg-team/pytorch_geometric#861 (comment)). We could possibly circumvent this problem by using a requirements.txt file (see: Cannot install torch-geometric via requirement.txt pyg-team/pytorch_geometric#861 (comment)). Let me know if you're up for giving this a try!

If we're already bumping versions, perhaps we should try to address #208 right away: Would you mind checking if you can run with the latest stable versions of the torch* packages we use? :)

[asogaard]

Actually, let's just approve and merge this puppy, then we can always address #208 later. 🥳

[asogaard]

Tag @MortenHolmRep for the above comment re: installing torch/geometric