how to apply drop_edge function to a SparseTensor.

[kou18n]

If the edge_index is a SparseTensor. How to apply the drop_edge function?

SparseTensor(row=tensor([     0,      0,      1,  ..., 373171, 373172, 373172], device='cuda:0'),
             col=tensor([     0,      2,      0,  ..., 373171, 373169, 373172], device='cuda:0'),
             size=(373173, 373173), nnz=1113071, density=0.00%)

[wsad1]

At the moment dropout_edge doesn't support SparseTensor as input, but that should be easy to support, will send a PR for that.
For now you can do that following.

from torch_geometric.utils import dropout_edge, to_edge_index
sparse_edge_index = SparseTensor(...)
edge_index, _ = to_edge_index(sparse_edge_index)
edge_index, _ = dropout_edge(edge_index, p=0.5)
sparse_edge_index = SparseTensor(row = edge_index[0], col = edge_index[1], sparse_sizes=(3, 3))

[kou18n]

At the moment dropout_edge doesn't support SparseTensor as input, but that should be easy to support, will send a PR for that. For now you can do that following.

from torch_geometric.utils import dropout_edge, to_edge_index
sparse_edge_index = SparseTensor(...)
edge_index, _ = to_edge_index(sparse_edge_index)
edge_index, _ = dropout_edge(edge_index, p=0.5)
sparse_edge_index = SparseTensor(row = edge_index[0], col = edge_index[1], sparse_sizes=(3, 3))

Thank you for your quickly reply! I will try your code!

[kou18n]

Is there any way to set the seed, and can reproduce the result of the dropout_edge function result?

[rusty1s]

I don't see a reason why you would want to change the graph size after applying dropout_edge. For changing the graph size in dropout_node, you can do

data.edge_index, edge_mask, node_mask = dropout_node(...)
data.x = data.x[node_mask]

[kou18n]

Thank you for reply.
I'm planning to implement the GraphCL model, and I need to generate two graph augmentation Aug1 and Aug2.
So I want to use dropout_node function to get one of the graph augmentation 'Aug1'. I think Aug1's node feature dim is differet with the orginal graph, maybe need to remove the dropped node feature from the orignal graph to get a new aug1 graph.

Original graph:
Data(y=[1], x=[1611, 5], adj_t=[1611, 1611, nnz=8759], num_nodes=1611)
Through dropout_node for example, 100 nodes have dropped.
Aug1 will become :
Data(y=[1], x=[1511, 5], adj_t=[.......], num_nodes=1511)
I'm not sure if my idea is correct.

[rusty1s]

I don't know the GraphCL paper, but what is the problem with that? This looks correct, right?

[kou18n]

Thank you for your today‘s answers.
#7583 problem solved.
About #1773, everything is done.
Currently, the code only support spares_tensor.

[rusty1s]

Will take a look, thanks!