search

rosinality / stylegan2-pytorch

Implementation of Analyzing and Improving the Image Quality of StyleGAN (StyleGAN 2) in PyTorch
MIT License
2.73k stars 619 forks source link

What's the meaning of augment_p? #108

Open cientgu opened 4 years ago

cientgu commented 4 years ago

It seems that the augment comes from "Training Generative Adversarial Networks with Limited Data", But I am confused what 'augment_p' and 'ada_target' and 'ada_length' means. Can I use these augmentation by only setting args.augment to True? Or do I need to change some other hyper parameters? Thank you so much!

rosinality commented 4 years ago

It is for adaptive augmentation scheduling for paper. If augment_p is not zero then augment will be applied with specified probabilities. If it is zero, then adaptive scheduling will be used. ada_target and ada_length is hyperparameter for adaptive scheduling, and it means each step of adaptive scheduling is adjusted so scheduling could make augmentation probabilities reach ada_target in the ada_length steps. You can find some details from the paper.

But I don't know my implementation is exactly matches with author's settings. You can try default hyperparameters with caution.

zrporz commented 9 months ago

I have the same question! And I can't find the augmentation details from the paper "Analyzing and Improving the Image Quality of StyleGAN" or "A Style-Based Generator Architecture for Generative Adversarial Networks". Does anyone have an idea?