torch_geometric.loader clusterData Issue with Partitioning

I’m looping through and trying to cluster large graphs into smaller partitions. When I use the clusterData function like so:

graph = pytorch_geometric.Data.data object 
graph2 = pytorch_geometric.Data.data object 


clusters1 = ClusterData(graph, 200)
clusters2 = ClusterData(graph2,200)

etc. (I’m doing much more than 2 graphs but it’s the same code repeated in a loop)

The code works for some of the graphs and I get the expected output:

Computing METIS partitioning... 
Done!

but then half way through I get this error:

Process finished with exit code -1073741819 (0xC0000005) I’ve gotten a similar partitioning error across both a Mac and a windows 10 laptop. If someone could help me out with this or at least point me to some resources where I might go about solving this issue it would be so helpful.

Environment: python=3.7.11 torch==1.9.0 torch_cluster== 1.5.9 torch_geometric== 2.0.1 torch_scatter== 2.0.8 torch_sparse== 0.6.12 torch_spline_conv== 1.2.1 torchmetrics== 0.5.1

[rusty1s]

I see. You can copy the ClusterData class and apply some modifications to it (adding self-loops, adding reverse edges), e.g.:

edge_index, _ = add_self_loops(data.edge_index)
edge_index, _ = to_undirected(data.edge_index)
adj = SparseTensor(
    row=edge_index[0], col=edge_index[1],
    value=torch.arange(edge_index.size(1), device=data.edge_index.device),    
    sparse_sizes=(N, N))

[licycle]

which jupyter kernerl version are you using?