Convert back from a dense batch to a sparse batch

[AnimeshSinha1309]

I need to use the pyg.utils.to_dense_batch function for getting dense_mincut_pooling to work.
Now, for the future layers, the tutorial on dense_mincut_pooling shows that we need to use DenseGraphConv layers.
However, there are some custom layers that I have implemented which do not have their Dense versions.
So is there some way, even if it is computationally inefficient, to undo the to_dense_batch function and get a sparse batch back.
I can't find a function called to_sparse_batch, is that by design?

[rusty1s]

You should be able to do

x = x.view(-1, x.size(-1)
edge_index, edge_weight = torch_geometric.utils.dense_to_sparse(adj)

As you said, this will return a lot of edges (due to how these dense pooling approaches work) which may be utterly inefficient. You may be able to reduce the number of edges by filtering them out based on edge_weight:

mask = edge_weight > 0.2
edge_index, edge_weight = edge_index[:, mask], edge_weight[mask]