eurunuela
commented
1 month ago I have tried adding the functionality on a local branch, but I keep getting this error:
File "/home/eneko/GitHub/tmima/tmima/training/model.py", line 108, in forward
x = self.down_convs[0](x, edge_index, edge_weight)
File "/home/eneko/GitHub/tmima/tmima/training/model.py", line 117, in torch_dynamo_resume_in_forward_at_108
edge_index, edge_weight = self.augment_adj(edge_index, edge_weight, x.size(0))
File "/home/eneko/GitHub/tmima/tmima/training/model.py", line 149, in augment_adj
edge_index, edge_weight = remove_self_loops(edge_index, edge_weight)
File "/home/eneko/GitHub/tmima/tmima/training/model.py", line 151, in torch_dynamo_resume_in_augment_adj_at_149
adj = to_torch_csr_tensor(edge_index, edge_weight, size=(num_nodes, num_nodes))
File "/home/eneko/GitHub/tmima/tmima/training/model.py", line 152, in torch_dynamo_resume_in_augment_adj_at_151
adj = (adj @ adj).to_sparse_coo()
RuntimeError: expand is unsupported for SparseCsr tensors
🚀 The feature, motivation and pitch
I noticed that the GraphUNet model doesn't make use of the edge attributes that I have calculated. I find it very odd that the model uses ones as the starting weights, as opposed to having the option to pass predefined weights to better inform the model on an initial pass.
Alternatives
No response
Additional context
No response