Some questions about networks_stylegan.py

[XiaoqiangZhou]

Thanks for your reimplementation!

1. In networks_stylegan.py, line 107 f = f[:, :, ::-1, ::-1], I think you should consider use f = torch.flip(f, [2, 3])
2. In networks_stylegan.py, line 247, x = self.noise(x, noise), I think you should consider add a condition judgement before this line, like if self.noise is not None:
3. In defination of forward function of the class GBlock, do you miss a 3x3 convolution between upsmaple and the first adaIn1?

[tomguluson92]

Thanks a lot!
It has been a few months that I haven't kept pace with progress in PyTorch, could you please explain is there a difference between f = f[:, :, ::-1, ::-1] and torch.flip(f, [2, 3])?
The advice 2, 3 looks good to me, you could pull request a version to improve it and i will have it look and merge if it works well.

^.^
Samuel

[XiaoqiangZhou]

Thanks for your reply.
For the first problem, if you use [::-1] index for a torch.tensor, you will get an error ValueError: negative step not yet supported, i.e., cannot flip a torch.tensor like numpy array.
Besides, I'm working on this project and it will be my pleasure to pull a request later.

[tomguluson92]

Thanks for your reply.
For the first problem, if you use [::-1] index for a torch.tensor, you will get an error ValueError: negative step not yet supported, i.e., cannot flip a torch.tensor like numpy array.
Besides, I'm working on this project and it will be my pleasure to pull a request later.

Thanks!

[ulucsahin]

Thanks for your reimplementation!

1. In networks_stylegan.py, line 107 f = f[:, :, ::-1, ::-1], I think you should consider use f = torch.flip(f, [2, 3])
2. In networks_stylegan.py, line 247, x = self.noise(x, noise), I think you should consider add a condition judgement before this line, like if self.noise is not None:
3. In defination of forward function of the class GBlock, do you miss a 3x3 convolution between upsmaple and the first adaIn1?

Did you find the answer for 3? I was thinking the same about missing a convolution between upsample and first adaln1 so I checked the issues and saw your question. I think the layer is missing as well but I haven't tested the code myself yet so I don't know if it works or not.