Selecting subset of edge indices

[ahmedramly]

I have a question about selecting a subset of edge indices. I have the following situation, I have 14 Nodes and the experiments include trying different combinations of edges and edge attributes so I want to do the following:

• Have the edge indices of a full graph which is of size torch.Size([2, 196])
• Have the edge attributes of a full graph which is of size torch.Size([196, N]) where N is the number of edge features
• Using a subset of edge indices of size torch.Size([2, m]) to define the graph and the edge attributes where m is the number of the edges in the subset.

It is a naive question but I am kinda stuck with it for a while. Thanks in advance

[wsad1]

One way to do this is as follows.

1. torch.cartesian_prod(torch.arange(196), torch.arange(196)).T create a fully connected graph.
2. use dropout_adj with training as true, to sample edges.

[ahmedramly]

I think I didn’t clarify my question in a correct way. The goal is not to drop some edges randomly with a probability. The problem is I don’t have the graph structure explicitly so I use different statistical and graph signal processing methods to learn the edges so in code it should be like this

G = Data(x=torch.Size([14, n_feats]), edge_index=torch.Size([2,196]), edge_attr=torch.Size([196,n_attrs]), y=torch.Size([1]))

G.subset_1 = torch.Size([2, 28])
G.subset_2 = torch.Size([2, 34])

subset_1 and subset_2 are already included in edge_index so the final goal is when training a network to define the edge_index and edge_attr in the forward As something like this

model(x = x, edge_index=subset_1, edge_attr=edge_attr[G.subset_1])

But it doesn’t work like this , I can’t use the subset_1 edge index as index to extract the corresponding attributes

Did I clarify the problem well?

[wsad1]

Initially you have a graph with say 14 nodes and 196 edges (fully connected). Then you predict which edges actually exist , and then run message passing using only the predicted edges. Is this an accurate summary. If so you could approach this 2 ways.

1. The edge classifier could return an edge mask. An edge mask would be a tensor.bool of shape num_edges (196 in your case). Using this you could do the following model(x = x, edge_index=edge_index[:,edge_mask], edge_attr=edge_attr[edge_mask]). This is a clean and efficient approach
2. Another approach is to modify your current code as follows:
   model(x = x, edge_index=subset_1, edge_attr=edge_attr[torch.all(torch.eq(edge_index, subset_1), dim=0)])

[rusty1s]

To chime in, another approach is to convert subset_1 into a one-dimensional vector which can be used to query edge_attr, e.g.:

idx_1 = subset_1[0] * 14 + subset_1[1]
edge_attr = edge_attr[idx_1]

Note that this approach only works in case edge_index and edge_attr encode a fully-connected graph though :)

[ahmedramly]

Thank you all for the help, I tried all of them and most convenient one is the last one by @rusty1s .
Also I should mention that the second solution by @wsad1 didn’t work with me, it gave me this error
RuntimeError: The size of tensor a (196) must match the size of tensor b (28) at non-singleton dimension 1
Thank you for the help again :)