RuntimeError when executing "test backward pass" block in main.ipynb

Question

RuntimeError when executing "test backward pass" block in main.ipynb

bluesquinn777 opened this issue 2 years ago · 7 comments

Hi, when I directly run your code in main.ipynb, other parts are correct, but I received "RuntimeError: grad_input must be contiguous"

I think this may be caused by loss_edges function, where y = y.permute(0, 3, 1, 2) which makes y not contiguous anymore? Thanks!

chaitjo commented 2 years ago

Great

Answer 1 · 2022-04-20T08:33:15.000Z

I had the same problem at runtime.

Answer 2 · 2022-04-20T08:39:51.000Z

Did you fix this discontiguous problem？

Answer 3 · 2022-04-20T10:16:34.000Z

Thanks for your interest - please make sure that you use the correct versions of Cuda and PyTorch as prescribed in the README.

Answer 4 · 2022-04-20T10:27:04.000Z

Thanks for your interest - please make sure that you use the correct versions of Cuda and PyTorch as prescribed in the README.

We ran our code on Ubuntu 16.04, using Python 3.6.7, PyTorch 0.4.1 and CUDA 9.0.

Thanks for your reply, I'm using windows, Python 3. CUDA 11.3, PyTorch 1.11.0, which may lead to this problem. Add .contiguous() after permute in the loss function can solve this problem.

Answer 5 · 2022-04-20T11:40:54.000Z

I configured it according to the README, and it is running successfully now, thanks to the author.

Answer 6 · 2023-03-17T08:33:22.000Z

.contiguous()

where do i need to add this tks ! @bluesquinn777