find the dropped edges when dropout_edge force_undirected=True

[cxw-droid]

Hi, how can I find the dropped edges after using dropout_edge? If force_undirected=False, it is easy:
dropped_edges = original_edge_index[:, ~edge_mask]
But if force_undirected=True, how can I find the dropped edges? Thanks.

[rusty1s]

I think you can still use edge_mask for this purpose, right? It is just that it only contains the upper triangle of edges that have been dropped.

[cxw-droid]

Thank you for your reply. If force_undirected=True, the returned edge_mask seems only masking the un-dropped nodes. So it seems not clear how to get the dropped nodes.

[rusty1s]

Ah, got it. This is indeed not really elegant (especially since edge_mask returns a LongTensor). Let me try to fix it.

[rusty1s]

Sorry for late reply. On a second look, this seems pretty difficult to implement. I suggest to get the dropped edges via

dropped_edge_mask = torch.ones(edge_index.size(1), dtype=torch.bool)
dropped_edge_mask[edge_mask] = False

[cxw-droid]

Thank you for your reply. The second line of code may not be correct if force_undirected=True. edge_mask only shows the edges in one direction. The edges in the opposition directions are also kept (not dropped).

[EdisonLeeeee]

edge_mask only shows the edges in one direction.

Yes this is intended because the opposite edges may not exist in the input edge_index. So how about using to_undirected after dropout_edge?

edge_index, mask = dropout_edge(edge_index)
edge_index, mask = to_undirected(edge_index, mask)