GraphMask Implementation for the main PyG library

[fork123aniket]

GraphMask Explainability technique is appearing in the torch_geometric.contrib.explain package. @rusty1s Any idea as to how to get this implementation to completion so that we can have it in the main PyG library??

To make the context more clear, the current version of GraphMask runs successfully for three different GNN layers: GCNConv, GATConv, and FastRGCNConv on three different GNN tasks: Node Classification, Graph Classification, and Link Prediction.

Thank you.

[rusty1s]

Sorry for my late reply here. Is there a problem with this being in contrib? I don't see this package as less valuable, we just prefer to keep complex external implementations there. If this is a problem, let me know and I will take a look to see what needs to be done to move this to torch_geometric.explain.

[rusty1s]

Super, will take a look.

[fork123aniket]

@rusty1s Thanks so much for merging the pull request yesterday. Really appreciate this move!! However, I do see the implementation is still in torch_geometric.contrib.explain. May I ask what are the expectations that the existing GraphMask Explainer implementation requires to meet in order to have it included in the main torch_geometric.explain package??

[rusty1s]

Hey @fork123aniket, yes I needed to move it back in order to see the changes/diff made in this PR. Otherwise, this was not possible. Next step would probably to add some more documentation to functions and add type hints to their arguments.

[fork123aniket]

Hi @rusty1s . As advised above, have included details in more functions and added datatype hints to the functions' arguments. Please consider taking a look at the PR #7503 and see if this GraphMaskExplainer implementation now can be included in the main library (torch_geometric.explain) or do let me know if something is still missing which needs to be included before driving this implementation from torch_geometric.contrib.explain to torch_geometric.explain. Thanks!!

[fork123aniket]

Hi @rusty1s. Now that we see PR #7504 has been merged, may I know what all expectations are left that need to be fulfilled that could aid in driving the implementation from torch_geometric.contrib.explain to torch_geometric.explain?? Could you shed some light on this, please??

[fork123aniket]

Hi @rusty1s, any update on this?? Can you guide me as to what all changes are yet to be made to make the transition of GraphMask Explainer from torch_geometric.contrib.explain to torch_geometric.explain?? Your advice concerning this transition may indeed help expedite this integration of GraphMask Explainer into the main library.

Thank you. Looking forward to hearing from you very soon.

[fork123aniket]

The purpose of using the LayerNorm module is to apply per-element normalization. Despite various attempts, couldn't find torch.linalg.norm's arguments' values in such a way that it provides output with shape (N, *) considering the same input shape. Rather, I did the following to have the output tensor with the same shape as that of the input's, yet achieved per-element normalization:-

norm = Sequential(LayerNorm(out.size(-1)).to(out.device), ReLU())
basis_messages = norm(out)

altered to

basis_messages = torch.nn.functional.layer_norm(out, (out.size(0), out.size(-1))).to(out.device).relu()

So, can we use torch.nn.functional.layer_norm instead of torch.linalg.norm, given both input (out) and output (basis_messages) tensors must have the same shape??

[rusty1s]

Yes, let's use torch.nn.functional.layer_norm instead, this would be great :)

[fork123aniket]

Done!! If no more changes are required, shall I tweak the code to ease the integration of the current GraphMask implementation into torch_geometric.explain from torch_geomtric.contrib.explain just by changing the paths in the code and attach other relevant code files (example.py and test.py) to the upcoming PR for final review?? What do you recommend??

[rusty1s]

Thanks. Can you send a PR with the corresponding changes? Let's do the final migration in a follow-up.

[fork123aniket]

As advised, have just opened a follow-up PR (#7779) for this final migration of the GraphMaskExplainer implementation. Please consider giving it a review.