cannot request out_edges() for empty node sets on cuda

[chwan1016]

🐛 Bug

I'm using the latest DGL version. The following code throws a DGLError: Check failed: dim != 0 (0 vs. 0) :

import dgl
import torch

g = dgl.DGLGraph(([0, 1], [1, 2])).to('cuda:0')
idx = torch.LongTensor([]).to('cuda:0')
g.out_edges(idx)

However, if the graph runs on CPU, the code works well. I'm not sure whether this is a bug.

[VoVAllen]

Empty tensor should not be allowed in out_edges function. Do you have any expected behavior here?

[chwan1016]

Actually, I just feel it's strange that out_edges has different behavior on CPU and GPU. It would be better if it returns an empty edge set.

[chwan1016]

A similar problem is found in out_degrees

[BarclayII]

Should be fixed in master. Will go into 0.5.2 release.