GNNExplainer for point cloud classification

[vctorwei]

Hi

May I ask if there is a way to apply GNNExplainer on point cloud classification model such as MPNN. (e.g. get importance of points in points cloud)

I learnt from this issue that we can get node importance from GNNExplainer, however, the graph generated from point cloud is always changing so I'm not sure if this is feasible.

By the way, I have another question, I cannot find the reference for the benchmark models of point clouds, MPNN and PointNet++ seem to have the same architecture with different kernels, may I ask what is the reference for this architecture (e.g. generating graph from point clouds before each graph convolutional layers)?

Regards

Victor

[rusty1s]

The recent changes in GNNExplainer (currently in master) allow you to disable the training of an edge mask via allow_edge_mask=False, and learn the importance of nodes (rather than the importance of features) via feat_mask_type="scalar". This might be helpful for your specific problem.

Yes, we use similar architectures to fairly compare different kernel formulations. Notably, the EdgeCNN will "learn" graph connectivity after each layer, while the others only operate on the initial one.

[rusty1s]

Thanks. I think the final issue is that the edge mask is applied (although untrainable) despite the allow_edge_mask option being set to False. We need to fix this in PyG directly. Are you interested in contributing this feature?

[vctorwei]

Thank you for the hint! Here is my understanding of this error, can you check if any of this is correct?
I got these two assumptions because when I remove the following steps, there will be no errors anymore.

        idx = fps(pos, batch, ratio=0.5)
        x, pos, batch = x[idx], pos[idx], batch[idx]
        radius = 0.4
        edge_index = radius_graph(pos, r=radius, batch=batch)
        pseudo = pos[edge_index[1]] - pos[edge_index[0]]

１.　Looks like it's the graph generation from pos steps between two conv layers that changed the node x and edge_index thus the edge_mask from the previous convolution does not match the node dimension?

If the problem above is correct then the solution would be that I'll need to pass the edge_mask of the newly generated edge_index all the way from the GNNExplainer function to the message_passing function?

２.　The graph generation step here using radius_graph is non-deterministic， which caused the same problem as this issue。

If this is the problem, I guess I'll just need to change the graph generation to a deterministc one? I learnt that radius_graph in CPU operation is deterministic, so I tried this which still does not work

        idx = fps(pos, batch, ratio=0.25, random_start=False)
        x, pos, batch = x[idx], pos[idx], batch[idx]
        radius = 0.3
        edge_index = radius_graph(pos.to('cpu'), r=radius, batch=batch.to('cpu'),  loop=False, max_num_neighbors=1000000)
        edge_index=edge_index.to('cpu')
        print(edge_index.size())
        pseudo = pos[edge_index[1]] - pos[edge_index[0]]

Any suggestions?

[rusty1s]

The problem is that edge_index changes during GNN execution as a result of pooling. Currently, GNNExplainer assumes fixed graph structure. This should be fixable by not adding the edge_mask attribute to each GNN operator in the first place (assumung allow_edge_mask=False), see https://github.com/pyg-team/pytorch_geometric/blob/master/torch_geometric/nn/models/gnn_explainer.py#L97-L107.

[vctorwei]

Since the edge_mask does not matter in this problem, I guess it's ok to directly comment outthe edge_mask related lines to avoid the assert errors? If the following lines are comment out, there will be no errors, I wonder if this will affect the results?

pytorch_geometric/torch_geometric/nn/conv/message_passing.py

Lines 326 to 327 in 682ab6c

	assert out.size(self.node_dim) == edge_mask.size(0)
	out = out * edge_mask.view([-1] + [1] * (out.dim() - 1))

[rusty1s]

Yes, that is one option, which does not affect the results. It sadly requires local modification to PyG, and a better option is to not set the edge_mask in the first-place in case allow_edge_mask=False.