Inconsistent number of neighbors being returned from knn_graph

[SergioRAgostinho]

Version: 1.4.5

I'm getting an inconsistent number of nearest neighbours being returned from from knn_graph (cuda). I'll try to provide the actual data later. Currently I'm invoking the method as

edge_index = knn_graph(P, k=5, batch=batch, loop=False)

P is of dim [2048, 64], composed of a single batch
batch is a vector of zeros size 2048
The flow is set to its default value source_to_target.

edge_index is showing the following values

# everything fine with this one. 5 neighbours per point
(Pdb) edge_index[1,:16]
tensor([0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 3], device='cuda:0')
# at the end there are problems. indexes 2046 and 2047 are showing 6 times neighbours instead of 5
(Pdb) edge_index[1,-16:]
tensor([2045, 2045, 2045, 2045, 2046, 2046, 2046, 2046, 2046, 2046, 2047, 2047,
        2047, 2047, 2047, 2047], device='cuda:0')

[SergioRAgostinho]

data.zip

Extract the archive and load the data with torch.load. It's a dictionary object containing the contents of P, batch and edge_index.

[rusty1s]

Thank you! I will have a look.

[rusty1s]

Ok, I identified the issue. All entries in P are equal, so removing self-loops leads to an undefined behaviour since the loop may or may not be included in the initial edge_index. I am not really sure if this needs to be fixed since this is an untypical use-case for knn_graph.

[SergioRAgostinho]

All entries in P are equal

Ok weird. If I'm interpreting what you wrote correctly, it means I missed a pretty naive mistake during debug. It wouldn't be the first time 🤷‍♂️. I'll recheck things.

[github-actions]

This issue had no activity for 6 months. It will be closed in 2 weeks unless there is some new activity. Is this issue already resolved?