conv2d_gradfix not supported on pytorch `1.10`

[rcharan]

This line is not forwards compatible with PyTorch 1.10 and the fallback leads to RuntimeError: derivative for aten::grid_sampler_2d_backward is not implemented.

I will follow up with a full stack trace and a PR to address this.

Edit: I see that this repository isn't accepting PRs, hopefully this will help anyone seeing this issue in the future.

Editing this line as well (which runs the same check) fixes this issue.

CUDA 11.4 with torch==1.10.0+cu111, tested on V100s and K80s

[rcharan]

(Related to #195)

[JavierCoronaA01023063]

TAHNKS

[snoop2head]

I created a pull request #197 for the fix! Hope it helps 🤗

[zhijiejia]

Make sure, the cuda is 11.1 or later, notice it is not the cuda that comes with pytorch, but the cuda of the machine itself, it is locate in /usr/local/cuda. It needs to be downloaded and installed manually.

The official also gave the answer in styleGAN3，can be see in https://github.com/NVlabs/stylegan3/blob/main/docs/troubleshooting.md

[Gword]

I created a pull request #197 for the fix! Hope it helps 🤗

Thank you! This works for me!

[bmquynhlinh]

This line is not forwards compatible with PyTorch 1.10 and the fallback leads to RuntimeError: derivative for aten::grid_sampler_2d_backward is not implemented.

I will follow up with a full stack trace and a PR to address this.

Edit: I see that this repository isn't accepting PRs, hopefully this will help anyone seeing this issue in the future.

Editing this line as well (which runs the same check) fixes this issue.

CUDA 11.4 with torch==1.10.0+cu111, tested on V100s and K80s

It works for me! Perfect! With RTX 3060, CUDA 11.5, and Torch 10.0.2+111

[EtDu]

I created a pull request #197 for the fix! Hope it helps hugs

Upon applying this patch, I'm getting the following:

RuntimeError: aten::grid_sampler_2d_backward() is missing value for argument 'output_mask'

pytorch                   1.11.0          py3.7_cuda11.3_cudnn8.2.0_0    pytorch
torchvision               0.12.0               py37_cu113    pytorch

[pedromartip]

I created a pull request #197 for the fix! Hope it helps hugs

Upon applying this patch, I'm getting the following:

RuntimeError: aten::grid_sampler_2d_backward() is missing value for argument 'output_mask'

pytorch                   1.11.0          py3.7_cuda11.3_cudnn8.2.0_0    pytorch
torchvision               0.12.0               py37_cu113    pytorch

You can fix this problem by modifying if any(torch.__version__.startswith(x) for x in ['1.7.', '1.8.', '1.9'] including the torch version you are using. This code is on grid_sample_gradfix.py.

[Ayesha-Imr]

I created a pull request #197 for the fix! Hope it helps hugs

Upon applying this patch, I'm getting the following:

RuntimeError: aten::grid_sampler_2d_backward() is missing value for argument 'output_mask'

pytorch                   1.11.0          py3.7_cuda11.3_cudnn8.2.0_0    pytorch
torchvision               0.12.0               py37_cu113    pytorch

You can fix this problem by modifying if any(torch.__version__.startswith(x) for x in ['1.7.', '1.8.', '1.9'] including the torch version you are using. This code is on grid_sample_gradfix.py.

The problem persists still, even after doing so. Any way of resolving this problem?

[pedromartip]

I created a pull request #197 for the fix! Hope it helps hugs

Upon applying this patch, I'm getting the following:

RuntimeError: aten::grid_sampler_2d_backward() is missing value for argument 'output_mask'

pytorch                   1.11.0          py3.7_cuda11.3_cudnn8.2.0_0    pytorch
torchvision               0.12.0               py37_cu113    pytorch

You can fix this problem by modifying if any(torch.__version__.startswith(x) for x in ['1.7.', '1.8.', '1.9'] including the torch version you are using. This code is on grid_sample_gradfix.py.

The problem persists still, even after doing so. Any way of resolving this problem?

Try to use the grid_sample_gradfix.py of the StyleGAN3. Just copy & paste the file.