it seems the codes are not working yet

[linghai06]

thanks for your sharing, but when I feed the trainer with a bunch of images from FFHQ(~1K), the loss stayed unchanged through each training iteration. any idea?

[plmsmile]

Thanks for the sharing~ I also met the problem. I trained 30k person images and generated 2 black images. I found the loss is NAN.

[linghai06]

Thanks for the sharing~ I also met the problem. I trained 30k person images and generated 2 black images. I found the loss is NAN.

you might get 0 std for the nan problem

[tomguluson92]

Thanks for your attention. This project is still under working, it’s a unfinished version. If you could help me fullfill this, it would be great!

[tomguluson92]

You can focus on Generator part.

[joe1chief]

Thanks for your attention. This project is still under working, it’s a unfinished version. If you could help me fullfill this, it would be great!

Is the tanh() necessary here? https://github.com/tomguluson92/StyleGAN2_PyTorch/blob/b953a8e4c3747989656e7c06be4fc650e4346e80/network/stylegan2.py#L1038