Pytorch Sparse tensor support: AntiSymmetricConv, CGConv, and TransformerConv
#6633
Conversation
Codecov Report
@@ Coverage Diff @@
## master #6633 +/- ##
==========================================
- Coverage 89.33% 87.73% -1.61%
==========================================
Files 422 422
Lines 22926 22887 -39
==========================================
- Hits 20482 20080 -402
- Misses 2444 2807 +363
📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
| out['edge_attr'] = None if values.dim() == 1 else values | ||
| if out.get('edge_type', None) is None: | ||
| out['edge_type'] = values | ||
| out['edge_type'] = None if values.dim() == 1 else values |
There was a problem hiding this comment.
Can you explain? Not sure I understand.
There was a problem hiding this comment.
Unlike torch_sparse.SparseTensor, edge_attr will never be None for PyTorch SparseTensor, which is not applicable for layers that accept edge_attr, e.g. TransformerConv
pytorch_geometric/torch_geometric/nn/conv/transformer_conv.py
Lines 220 to 221 in a5e5a48
There was a problem hiding this comment.
Ah, I see. This is tricky. I reverted the edge_type chance back since the values are always expected to be one-dimensional, but otherwise, I think the fix looks good to me.
A minor change on
nn.MessagePassinglayer, whereedge_attrandedge_typeareNoneby default if the sparse adjacency matrix is 2-dimensional.