Split batched big graph into original sub-graph or masking

[JueXiuHuang]

Hi,
currently I'm working on a Graph2Text task, and I'm trying to do graph-text alignment in my model.
But DataLoader in PyG will stack the batched sub-graphs into a big graph, how can I split them into original sub-graph or using a mask for alignment?

Thank!

[fengsxy]

I have the the same question,i will prepare three graph together,the model need to handle the three graph.At first , i try to concate them.And then use index to split them.

But then because of the stack charaters,the index is over it.
for example my x [9,484]
but edge_index is over 3000
so i want find a way to solve it !

[rusty1s]

Yes, you will need to take care of decrementing edge_index for each individual example. You can use SparseTensor for this (which does it automatically):

from torch_sparse import SparseTensor
adj = SparseTensor.from_edge_index(edge_index, edge_attr)
adj = adj[start_idx:end_idx, start_idx:end_idx]
(row, col), edge_attr = adj.coo()
edge_index = torch.stack([row, col], dim=0)

Does that answer your question?

[fengsxy]

yeah,Thank u for your answer, i think that may be ok. This time I use the information which is edge length i stoage in dataset to recompute the edge_index,and that's have already been ok.But i still have one question ,if i have three pairs of edge_index,edge_attr,can i use anything
to make three pair of edge indexs stack independtly and automatically.Like the one.Really thanks for your answer.

[rusty1s]

You can also use SparseTensor for this:

from torch_sparse import cat, SparseTensor

adjs = [SparseTensor.from_edge_index(edge_index) for edge_index in edge_indices]
adj = cat(adjs, dim=(0, 1))  # diagonal stacking

[rusty1s]

@JueXiuHuang Can you clarify on what you want to do exactly? In general, you can use batch to index select the components of an individual graph. Furthermore, the to_dense_batch function will convert your node features to shape [batch_size, num_nodes, num_features]. Does that help in your case?

[JueXiuHuang]

Thanks! This information helps a lot!