On Data Augmentation for GAN Training

Recent successes in Generative Adversarial Networks (GAN) have affirmed the importance of using more data GAN training. Yet it is expensive to collect many domains such as medical applications. Data Augmentation (DA) has been applied these In this work, we first argue that classical DA approach could mislead generator learn distribution augmented data, which be different from original data. We then propose a principled framework, termed Optimized for (DAG), enable use training improve learning distribution. provide theoretical analysis show our proposed DAG aligns with minimizing Jensen-Shannon (JS) divergence between and model Importantly, effectively leverages discriminator generator. conduct experiments apply models: unconditional GAN, conditional self-supervised CycleGAN datasets natural images images. The results achieves consistent considerable improvements across models. Furthermore, when used some models, system establishes state-of-the-art Frechet Inception Distance (FID) scores. Our code available.