Pyg explaination for edge classification model.

[arthurserres]

Hello all, I recently discovered the Explain module of torch geometric. I want to use it to explain an edge classification model. Unfortunately, I have two main issues:

1 - Is this framework supported by explain, maybe can you tell me where I can find an example for an edge classification model ?

2 - My graph is huge and I usually need to use loaders for training and evaluation (Because of memory limit). Some tests with explaination made my kernel crash, is there a way to use loaders, or is it a problem that you know how to adress ?

To be 100% clear, here is the project I try to make explainations about: Multi-GNN

Many thanks by advance.

Best,

Arthur

[xnuohz]

1. For GNN Explainer and Link Prediction, you may check this one: https://github.com/pyg-team/pytorch_geometric/blob/master/examples/explain/gnn_explainer_link_pred.py
2. For DataLoader, you can try here: https://pytorch-geometric.readthedocs.io/en/latest/modules/loader.html

[arthurserres]

1 - I was refering ot edge binary classification, not link prediction, any guess for that ?

2 - I was refering to loaders for the explain class, namely how to use loaders with explaination. My datas are too large to be charged in one time on the computer. Any idea about how to use loaders with explain. Or simply to manage memory issues with explain ?

Thanks by advance.

Arthur

[rusty1s]

1. Yes, both edge classification and link prediction is supported. In general, any task is supported as long as the output can be mapped to a given ground-truth.
2. This should work as well. You should be able to iterate over your graph and explain it in batches.

[arthurserres]

Thank you very much for the clarification, that is very nice of you. By any chance, do you have example of one of or both cases ?

[rusty1s]

This example here shows how to use Explainer in edge regression tasks: https://github.com/pyg-team/pytorch_geometric/blob/master/examples/explain/captum_explainer_hetero_link.py. We are currently missing an example within data loaders, but it really just comes down to calling explainer in each mini-batch iteration.

[arthurserres]

Thanks a lot for this answer. I will have a deeper look in it now. I will close the discussion and get back to you if needed !

Thanks a lot for your job :)

[arthurserres]

I looked again at this loader question and would like to check with you what should be done. What I understand from your answer is that we simply need to run something like

for batch in tqdm.tqdm(loader):  
    with torch.no_grad():
                batch.to(device)
                explanation = explainer(
                batch.x,
                batch.edge_index,
                index=index,
                edge_attr=batch.edge_attr
                )

and then at the end of the process, explaination will contain the explaination for the full graph, is that correct ?