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?

To give images of our own image and test it?

It seems that styleGAN only needs one dataset, unlike cycleGAN. But the offical video does show that it could do something like style mixing or transfer.
How to do that with styleGAN?

Thanks for your consideration and look forward to your early reply!

Thanks for sharing the code.

As shown in the line below. Why do you add x when performing the instance normalization?

x = x * (style[:, 0] + 1.) + style[:, 1]

All the best,
Xianxu

Hi~

Very thanksful for your pytorch code sharing. I am very interested in the performance of the this styleGAN in pytorch. Have you test it in some metric, like FID, or check the generated results in the middle phase of training?

thanks,

EK

Hi,
Thanks a lot for your pytorch version code.
I am training my own dataset with this model. But, the intermediate file till 80 epoch is still almost blank.
since my dataset has a lot of white background images, I am wondering if it is the main reason.

Thanks for the good work. But I can't find where you implemented the progressive growing, which is one of the essential part of StyleGAN. It seems in the G_synthesis module in networks_stylegan.py, only the 'fixed' mode is supported. Does it mean the whole training process isn't conducted in the progressive growing fashion when I do "python train_stylegan.py"? And do you intend to support this feature in the future?

I am not entirely sure, but if you're creating the codebase from scratch (from the paper) then would the artifacts generated then be a property of the licensor/owner of the intellectual property? If not then what is the commercial usage restriction on it?

karras2019stylegan-ffhq-1024x1024.pkl
whether the above file contains training dat info or the dataset info?
or the both on the ffhq pkl file
reply me once u go through this!

First, thanks for your great work！
And can you provide some results from the trained model?

Hello,
First of all thank you for the code in pytorch. I am currently trying to run your implementation without any change except for the dataset. It seems that from the beginning itself the Generator Loss and the Discriminator Loss is saturating at 0.6 and 1.3 respectively. After a few epochs they are generating nan values. Could you please help me with this.
Thank you

is these two parameters take the image input from that .pkl file?

We test this code on the FFHQ dataset, It doesn't work. There must be something wrong with this code, but we can't find it.

Hi, thanks for your reimplementation. It is so elegant.

But I found that you have a LeakyReLU layer after dense1, which does not exist in the tensorflow implementation.

https://github.com/NVlabs/stylegan/blob/66813a32aac5045fcde72751522a0c0ba963f6f2/training/networks_stylegan.py#L617

I think that adding activation after the last fully connected layer is not a good idea.

Hi~ thank you for converting styleGAN to its pytorch version, which is pretty meaningful. I've noticed that there are some people who can not get resonable training results. Unfortunately, i am one of them. I've only got the results below:

The training set i am using have 30k images, so i don't think that is caused by the number of training set images. Hope that you can help me with it~

What is normalize module right after latent z input?

Thanks for your code！
I run the program with default parameters and images in star folder , but only the noise image is generated after 300 epochs.
Is there a problem with my operation?

I have a DELL XPS with an nVidia 1080 w 8GB. My hope was to run StyleGAN on this rather than needing to jump to the cloud. Possible?

I disable blur2d by commenting out the lines including it

Getting this error:

:\Users\Powerpop\Anaconda3\lib\site-packages\skimage\transform\_warps.py:110: UserWarning: Anti-aliasing will be enabled by default in skimage 0.15 to avoid aliasing artifacts when down-sampling images.
  warn("Anti-aliasing will be enabled by default in skimage 0.15 to "
Traceback (most recent call last):
  File "train_stylegan.py", line 164, in <module>
    main(opts)
  File "train_stylegan.py", line 98, in main
    real_logit = D(real_img)
  File "C:\Users\Powerpop\Anaconda3\lib\site-packages\torch\nn\modules\module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "C:\Users\Powerpop\Desktop\StyleGAN\networks_stylegan.py", line 651, in forward
    x = F.leaky_relu(self.conv6(x), 0.2, inplace=True)
  File "C:\Users\Powerpop\Anaconda3\lib\site-packages\torch\nn\modules\module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "C:\Users\Powerpop\Anaconda3\lib\site-packages\torch\nn\modules\conv.py", line 338, in forward
    self.padding, self.dilation, self.groups)
RuntimeError: CUDA out of memory. Tried to allocate 4.00 GiB (GPU 0; 8.00 GiB total capacity; 4.19 GiB already allocated; 2.06 GiB free; 8.04 MiB cached)