Get wrong results when running random walk on gpu

[rusty1s]

Get wrong results when running random walk on gpu

[rusty1s]

Thanks for reporting. I will look into it. I guess python setup.py test also fails for you?

[EricGz]

Thank you for looking into it! You are right. The test failed.

[rusty1s]

Do all GPU tests fail?

[EricGz]

Yes, I think so. 55 failed and 56 passed. All the failed ones are GPU tests.

[rusty1s]

Ok, so this is not a problem with the random walk function, but the installation of torch-cluster. Can you post the log of

rm -rf build && python setup.py install

[EricGz]

Here's the log. log.txt

[EricGz]

It seems the whole installation goes fine. However, I still get wrong results running random walk on GPU. Do you have any idea what went wrong?

[rusty1s]

Unfortunately no :( Logs look okay to me. Maybe you have multiple versions installed where one installation failed? You can try removing torch-cluster repeatedly and install again.

[EricGz]

Hi, @rusty1s, thanks for your timely reply. I tried your suggestion, but the problem is still unsolved.

I tried some other tests and got more information about this error. When I used the GPU version of scatter_max and scatter_min in package torch_scatter, I met this error again, and the interesting thing is that the GPU version of functions' like scatter_add and scatter_mean worked fine.

Maybe there's something common about scatter_max and random_walk that caused the error?

P.S. Here's the test results of scatter_max and scatter_add

import torch
from torch_scatter import *

# device = 'cpu'
device = 'cuda:1'

src = torch.tensor([[1., 1.], [1., 1.], [4., 2.], [2., 4.]]).to(device)
index = torch.tensor([0, 0, 1, 1]).to(device)
index = index.view(-1,1).repeat(1,src.size()[1])

res1, _ = scatter_max(src, index, dim=0, fill_value=1.)
res2 = scatter_add(src, index, dim=0, dim_size=2, fill_value=0.)

print(res1)
print(res2)

The results are

tensor([[1., 1.],
        [1., 1.]], device='cuda:1')
tensor([[2., 2.],
        [6., 6.]], device='cuda:1')

I tried to debug it and I found that line 13 func(src, index, out, arg, dim) of max.py did not change the variable out at all. Do you have any clue about what caused the problem?

[rusty1s]

Yeah, those are the functions that call our own kernel implementations. It seems that there is something wrong with you GPU setup in conjunction with the provided cuda code.

[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?