Closed WPettersson closed 1 year ago
Thanks @WPettersson. PyTorch Geometric and CUDA have only specific compatibilites as shown here https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html#quick-start
We could add a similar table in the README to help users with installation on GPU machines.
An even easier solution would be to simply link to those instructions with a sentence like "GNS uses pytorch geometric and CUDA. These packages have specific requirements, please see insert_link_here
for details on installing pytorch-geometric."
It's up to you exactly how you phrase it, but I would give links for installing pytorch, as well as pytorch-{geometric,sparse,scatter,cluster} as these do have specific CUDA requirements (mostly, matching pytorch versions and cuda versions). You can then follow this with a sentence like "After installing the above, the remaining requirements can be installed with pip install -r requirements.txt
".
Nice to see that you opened issues here @WPettersson. I just add a link to the review here so we keep track: https://github.com/openjournals/joss-reviews/issues/5025
Your README file includes details on how to install/use gns on TACC, but doesn't have information for people who won't be using TACC. It might be very hard to write a singular script to handle all cases, but it would be good to at least have a list of required packages (preferably with a link on where to find them) and compatible version numbers. For instance, the
build_venv.sh
file pulls in cuda 11.3 - will other versions work? It's okay to say "we don't know" if you don't know, but it's helpful to also point out that it is meant to work with 11.3.