Some usability updates to Explanation

[BlazStojanovic]

Motivation

When evaluating and visualizing explanations it will be very convenient to obtain the explanation and explanation complement subgraphs directly from the Explanation objects. Given that the class holds information of the original graph being explained this shouldn’t be difficult to achieve. I propose we add two methods to the Explanation class:

1. get_explanation_subgraph returns the explanation subgraph $G_S$,
2. get_complement_subgraph returns the complement of the explanation subgraph $G_{C \backslash S}$

I’ve submitted this draft PR to get opinions on this, both implementation-wise and if we should do this in the first place.

Alternatives

1. Delegate this to other parts of code, I.e. whenewer this is needed (e.g. when evaluation explanations or visualizing them)
2. Implement the logic in util.py for general Data object, e.g. a mask_graph method, which can then be reused here or elsewhere.

[codecov]

Codecov Report

Merging #6054 (a9856f0) into master (6aa454a) will increase coverage by 0.02%.
The diff coverage is 100.00%.

❗ Current head a9856f0 differs from pull request most recent head 7320580. Consider uploading reports for the commit 7320580 to get more accurate results

@@            Coverage Diff             @@
##           master    #6054      +/-   ##
==========================================
+ Coverage   84.27%   84.29%   +0.02%     
==========================================
  Files         362      362              
  Lines       20456    20488      +32     
==========================================
+ Hits        17239    17271      +32     
  Misses       3217     3217              

Impacted Files	Coverage Δ
torch_geometric/explain/explanations.py	100.00% <100.00%> (ø)

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[BlazStojanovic]

@rusty1s and @dufourc1, while working on this PR, I’ve encountered the following behaviour. If you create a partially empty Explanation, e.g.

node_mask = torch.tensor([False, True])
explanation = Explanation(node_mask=node_mask)

then trying to acces either of the other masks results in the following AttributeError :

explanation.edge_mask (or node_feat_mask, edge_feat_mask)
...
AttributeError: 'GlobalStorage' object has no attribute 'edge_mask'

Is this intended behaviour?

[rusty1s]

I guess it is. You can access them via data.get('edge_feat_mask') if you want to access potentially non-existing attributes. Would that work for you?

[dufourc1]

I would have expected that to return None, even if it does not change a lot given the workaround rusty1s just gave 👍.