Version dependence of pytorch

[haotong-yang]

Hi! This work is excellent so I want to reproduce the codes as soon as possible. However, when I try to reproduce the codes, I find there are some bugs such as the parameter mismatch when loading checkpoint:

checkpoint['model']['graph'] is expected to be a tensor or tensor-like but get a graph.Graph

After trying, I thought maybe it was related to the version of PyTorch. After rolling back the PyTorch version from 1.12.1 to 1.10.1 (and reinstalling the torch-scatter), the bug is solved. I believe if you are willing to test this issue and give some instructions, it can help more people who love this job like me.

[KiddoZhu]

Hi! Thanks for pointing this out. This is generally a compatibility issue between TorchDrug and PyTorch, since TorchDrug overloads a few PyTorch modules. We will test TorchDrug on newer PyTorch versions and fix that.