pytorch_geometric: Segmentation Fault in Forward Loop of Edge Conv

Hey there!

I’m testing out edge conv and am running into some issues. I’m getting a segmentation fault during the knn_graph generation:

line 13: 14361 Segmentation fault      (core dumped) CUDA_VISIBLE_DEVICES="$gpuNum" python ...

Here’s output using pysnooper, I traced the error to this location:

Starting var:.. batch = None
Starting var:.. pos = tensor([[-3.1472e-01, -7.1309e-01, -1.5181e-01,  1.3493e+00,  1.0879e+00,          4.9691e-01, -1.54...
Starting var:.. self = Net(  (conv1): EdgeConv(nn=Sequential(    (0): Linear(in_features=50, out_features=64, bias=True)   ...
17:39:03.629809 call        41  def forward(self, pos, batch):
17:39:03.680347 line        42          edge_index = knn_graph(pos, k=20, batch=batch)
~

Last line is where error is happening.

As more context, I’m generating my point cloud using a CNN, batch and positions are shown above, cannot seem to make it through the generation of the edge index.

Can you please help me out here?