Different tensor indexing results on different machines

1

Hi, I have two different machines both Ubuntu 18.04 with torch 1.9.0 and 1.9.0+cu102. When I try the followings:

V = torch.tensor([[1,2,3],[4,5,6]])
idx = torch.tensor([0])

Now if I try:

print(V[idx])

I ll get different outputs:

tensor([[1, 2, 3]]) # for torch 1.9.0
tensor([1, 2, 3]) # for torch 1.9.0+cu102

Do you have any idea?

2

Assuming the difference between both installations would be a CPU-only vs. CUDA10.2 install, I cannot reproduce the issue and both pip wheels give tensor([[1, 2, 3]]) as the output using your code snippet.

3

@ptrblck this is very strange because the torch 1.9.0 is definitely using GPU.

4

Yes, that’s indeed strange, as I’ve installed the CPU and CUDA versions of 1.9.0 for this particular test and wasn’t able to reproduce it.

5

The problem is my own code which hooks torch and somehow affects the __getitem__ functionality of torch.