Issue in SplineCNN while clustering

[shreygadiya]

Hi,
I was using the SplineCNN with InMemoryDataset for one of my own datasets. I am getting the following errror in training while using graclus:

THCudaCheck FAIL file=aten/THC/generic/THCDegree.cuh line=23 error=11 : invalid argument
Traceback (most recent call last):
File "BACH.py", line 96, in
train(epoch)
File "BACH.py", line 75, in train
out = model(data)
File "/home/Drive2/shrey/shrey/venvSpline/local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "BACH.py", line 40, in forward
cluster = graclus(data.edge_index, weight, data.x.size(0))
File "/home/Drive2/shrey/shrey/venvSpline/local/lib/python2.7/site-packages/torch_geometric/nn/pool/graclus.py", line 6, in graclus
return graclus_cluster(row, col, weight, num_nodes)
File "/home/Drive2/shrey/shrey/venvSpline/local/lib/python2.7/site-packages/torch_cluster/graclus.py", line 34, in graclus_cluster
graclus(cluster, row, col, weight)
File "/home/Drive2/shrey/shrey/venvSpline/local/lib/python2.7/site-packages/torch_cluster/utils/ffi.py", line 12, in graclus
func(self, row, col) if weight is None else func(self, row, col, weight)
File "/home/Drive2/shrey/shrey/venvSpline/local/lib/python2.7/site-packages/torch/utils/ffi/init.py", line 202, in safe_call
result = torch._C._safe_call(*args, **kwargs)
torch.FatalError: cuda runtime error (11) : invalid argument at aten/THC/generic/THCDegree.cuh:23

The error pops up for one of the graph everytime (I am currently using batch size = 1 with no shuffling). Note that the no of nodes in each graph is same.

Any help in this regard would be appreciated. Thanks

[rusty1s]

Mh I guess I need more information to help. The error occurs always on the same graph?
Does the graph

• has no edges?
• contains self-loops?
• contains isolated nodes?

[shreygadiya]

Hi,
The graph has following properties:
Batch(batch=[9667], edge_attr=[1120425, 2], edge_index=[2, 1120425], pos=[9667, 2], x=[9667, 64], y=[1])

Since you ask, is not having any self-loops or isolated nodes a necessity? Many of the other graphs are rather disconnected at multiple places. And each node has a self loop in all the graphs.

Thanks.

[rusty1s]

Self-loops are removed internally before clustering is computed, so this should be no problem.
Isolated nodes shouldn't be a problem on GPU (the CPU version has an issue there).

Please check if edge_index.max().item() + 1 is less than or equal to 9667.

[rusty1s]

Any updates?

[shreygadiya]

Hi. I discarded the graphs that were giving me issue and training on the rest of the network. However, I stopped working on the project a while ago, so I don't know yet what the exact issue is with those graphs. I'll try to figure it out and get back to you in a week or two.

[shreygadiya]

Hi,
I wanted to inform that currently I am no longer working on this project. If you wish you can close this issue.
Thanks.