Class Balanced Sampling for the Training in GANs

Abstract

Recently Top-k fake sample selection has been introduced to provide better gradients for training Generative Adversarial Networks. Since the method does not guarantee class balance of selected samples in class conditional GANs, certain classes can be completely ignored in the training. In this work, we propose class standardized critic score based sample selection which enables class balanced sample selection. Our method achieves improved FID score and Intra-FID score compared to prior Top-k selection.