Getting transpose of edge features

[PabloVD]

I generate a graph using for instance torch_cluster.radius_graph, so the edges have shape $(2, N_e)$. Then I build edge features $e_{ij}$ between nodes nodes $i$ and $j$. with shape $(N_e,C)$, being $C$ the number of channels. Note that in general $e_{ij}$ is different from $e_{ji}$.

Given the edge features $e_{ij}$, how can I get the transpose edge feature, $e_{ji}$? Is there any efficient way of doing that in pytorch geometric?

[wsad1]

I am a little confused by the question. What do you mean by transpose of edge features? Lets say your graph has a single edge [[1,2]]
and the feature vector for that edge is lets say [0.1, 0.2, 0.3]. The edge transpose would be [[2,1]], what would the transpose of the features be? Do let me know if I misunderstood something.

[PabloVD]

Many thanks for your answer. Could you please explain how you define indexes i and j above? Are equivalent to row, col = edge_index[0], edge_index[1]?

[PabloVD]

That works! I managed to get the transpose tensor edge_attr_trans from a given tensor of edge features edge_attr. Not very efficient but works, many thanks @rusty1s !

from torch_sparse import SparseTensor
adj = SparseTensor.from_edge_index(edge_index, edge_attr)

edge_attr_trans = torch.zeros_like(edge_attr, device=device)

for edge in range(edge_index.shape[1]):
    i, j = edge_index[0,edge].item(), edge_index[1,edge].item()
    e_ji = adj[j, i].to_dense()
    edge_attr_trans[edge] = e_ji

[rusty1s]

Nice, I think you can also use j and i as vectors, which avoids the need of the for-loop.

[PabloVD]

Not sure how. If I try

i, j = edge_index[0], edge_index[1]
e_ji = adj[j, i].to_dense()

then, the resulting array has shape $(N_e, N_e, C)$ rather than the expected $(N_e, C)$ (being $N_e$ the number of edges and $C$ the number of channels).

[rusty1s]

Ah, you are right. Sorry about that. It looks like we really don't support this yet efficiently in SparseTensor :(