About celeba-hq.yaml
Hpjhpjhs opened this issue · comments
Thanks for your great works. I have some questions about the training configuration. There are differences between celeba-hq.yaml and celeba-hq_256.yaml in configs folder, such as the normalization in different channels. What's the reason here? If I try to train the model based on the size of 512 or 1024, how should I set? It is appreciated if I can receive your reple
Actually, the normalization methods are the same among different configs, while for different modules, they are different (AdaIN for translator, None for discriminator and extractor, IN for encoder and decoder). Non-normalization is discriminator is from most recent GANs like stylegan or starganv2. AdaIN is to make the output condition on the style, you can change this to demodulation following stylegan2. Using IN in encoder and decoder is the most uncertain one, you can try different normalization methods here (but I found that non-normalization will make the training unstable).
The difference between these two configs is the different channel numbers in the discriminator and extractor, which I have to decrease since the limited GPU memory (2x 1080Ti).
For high-resolution training, please kindly refer to this issue (https://github.com/imlixinyang/HiSD/issues/12), where I clarify why I haven't released a 512 or higher resolution checkpoint and some possible solutions.
@imlixinyang Thanks for your detailed explanation. I also tried some other methods like pixel2style2pixel to do image transfer, which can manipulate 1024 resolution of images, but most of them are strongly dependent on the encoded latent code。 So the inspiration of this work is amazing. Thanks for your great work again. By the way, I have a relative research background and strong interest in image translation research, So I want to ask whether the research of image-to-image translation is led by Pr. Ji or any other people in MAC lab?