T035 Graph Neural Networks: Package Installation and use of GPU

[janash]

Thanks for the wonderful set of tutorials! I was just going through the tutorial on Graph Neural Networks yesterday and I noticed a few things about the code:

(1) It checks for platform then installs pyg (pytorch geometric) using mamba if on Linux or macOS. This seems undesirable to me as it could unknowingly alter the user's environment and will also fail if the user has a different package manager. Perhaps there could be a note on how to install the required packages, or if this is for CI it could be handled a different way.

(2) The tutorial does not check for device or send data to GPU if available. This makes training the GNNs very slow. When completing the tutorial, I made this change.

If you'd like, I can submit a PR for (2), and I'd be happy to brainstorm on (1) and submit a PR there as well.

Thanks!

[mbackenkoehler]

Thanks for the feedback! Those are excellent points!

We currently are reworking the dependency management to split up environments. The current plan is to have one environment per talktorial. A startup script should install and run a given notebook within that environment. That's WIP on branch pip_sepenv. We would welcome any thoughts and feedback on that or other ideas!