How to get edges between clusters using "ClusterData"?

[ObsisMc]

Are there any methods can I use to get edges between different clusters after using ClusterData

[wsad1]

That class needs better documentation.
ClusterData contains perm and partptr. perm is the new permutation of nodes and partptr is of shape num_parts+1 and assigns node slices to a partition. You should be able to use these two to find edges that go across clusters.

[ObsisMc]

For example

Triangles and circles are nodes, and they are in different clusters. Green lines are edges in different clusters and red lines are edges connecting the two clusters.

Then, I want to get red lines.

[rusty1s]

If you know the cluster, you can do something like adj[partptr[i]:partptr[i+1], partptr[j]:partptr[j + 1]].

[ObsisMc]

Sorry but I cannot understand the adj and don't know how to use it.
In class ClusterData, adj is defined and operated here:

            adj = SparseTensor(
                row=data.edge_index[0], col=data.edge_index[1],
                value=torch.arange(E, device=data.edge_index.device),
                sparse_sizes=(N, N))
            adj, partptr, perm = adj.partition(num_parts, recursive)

Do meanings of row and col attributes in adj change after partition? It seems that the adj isn't in COO format since I find it contains some edges which don't exist in original graph.

[rusty1s]

Yes, row and col are permuted after the adj.partition call, so the indexing change such that the first n rows and n cols refer to the indices of the first partition. You can use the indexing logic as described above to get inter- and intra cluster connectivity.

[ObsisMc]

Thank you very much, I get it now. After adj[partptr[i]:partptr[i+1], partptr[j]:partptr[j + 1]], I can use value attribute to find the edges in my orginal graph.