search

Forschungsprojekt-Informatik / CompGCN

ICLR 2020: Composition-Based Multi-Relational Graph Convolutional Networks
Apache License 2.0
0 stars 0 forks source link

BUG: Anything using the helper.py with torch 1.13.1 fails #6

Closed A7exSchin closed 1 year ago

A7exSchin commented 1 year ago

Anything thats tries to use the helper.py functions ccorr, conj, ... with torch 1.13.1 instead of 1.4 fails with different errors, as upgraded torch has moved functions.

1Maxnet1 commented 1 year ago

Could you give an example command and report what happens when you execute it ?

JWG-1 commented 1 year ago

fixed, problem appeared because of version change of pytorch, the problem appeared when using the circular correlation

1Maxnet1 commented 1 year ago

fixed, problem appeared because of version change of pytorch, the problem appeared when using the circular correlation

Could you also describe how you fixed it? I couldn't find a related PR so it might make sense to keep the issue open until the solution is available in development

JWG-1 commented 1 year ago

the problem was in that in the newer version of pytorch the implementation of rfft and irfft are in torch.fft.xxx and that they changed outputted to a complex Tensor and not a 2 dimensional tensor. The solution was to simpley change the implentation of the helper functions conj(), com_mul(), cconv() and corr(). I committed the solution already to the development branch.

1Maxnet1 commented 1 year ago

the problem was in that in the newer version of pytorch the implementation of rfft and irfft are in torch.fft.xxx and that they changed outputted to a complex Tensor and not a 2 dimensional tensor. The solution was to simpley change the implentation of the helper functions conj(), com_mul(), cconv() and corr(). I committed the solution already to the development branch.

Ok then. As we defined in https://github.com/Forschungsprojekt-Informatik/CompGCN/wiki/Branching-model please do not commit or push directly to the development branch.