Open fierval opened 2 years ago
The issue is that there is no edge type pointing to category
. Upgrading to torch-geometric==2.0.4
should at least warn you about this.
Figured as much but thought code like None.relu()
was worth reporting.
Yes, definitely :) I think the current workaround of warning the user is okay, but I agree that in your case it should definitely crash prior to model execution. I need to look into torch.fx
to see if there is some way to check for this.
So every node in the heterogeneous graph should have an edge pointing to it? How should we resolve this problem? I tried to change the graph to Undirected but it didn't work. Should we add reverse edge type to our data by ourselves?
Yes, otherwise certain node types will not get properly updated during message passing. The ToUndirected
transform should take care of that. Let me know if you encounter any issues wit that.
🐛 Describe the bug
Converting the following simple model with
to_hetero()
:With metadata:
(['category', 'vertex'], [('category', 'assign', 'vertex'), ('vertex', 'groups', 'vertex'), ('vertex', 'continues', 'vertex'), ('vertex', 'keyof', 'vertex')])
Yields the following:
Then relu_category = None.relu() causes it to crash.
Environment
conda
,pip
, source): piptorch-scatter
): all