You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I am getting an autograd error when calling backward on data that has been fed through a simple graph conv network using a SparseTensor adjacency matrix. If I instead feed the original edge_index into the network, then autograd works fine.
ArchieGertsman
changed the title
Autograd error after successful forward pass with SparseTensor adjacency mat instead of edge_index
Autograd error after conv with sparse adjacency mat instead of edge_indexJan 9, 2023
Thanks @rusty1s, this works! I was following Memory-Efficient Aggregations and was unaware of the sparse_sizes argument. Are there docs for torch_sparse anywhere? I can't seem to find any.
馃悰 Describe the bug
I am getting an autograd error when calling
backward
on data that has been fed through a simple graph conv network using aSparseTensor
adjacency matrix. If I instead feed the originaledge_index
into the network, then autograd works fine.out:
Environment
conda
torch_sparse
version: 0.6.16The text was updated successfully, but these errors were encountered: