eduzhai > Applied Sciences > Engineering >

On Data Augmentation for GAN Training

  • Save

... pages left unread,continue reading

Document pages: 16 pages

Abstract: Recent successes in Generative Adversarial Networks (GAN) have affirmed theimportance of using more data in GAN training. Yet it is expensive to collectdata in many domains such as medical applications. Data Augmentation (DA) hasbeen applied in these applications. In this work, we first argue that theclassical DA approach could mislead the generator to learn the distribution ofthe augmented data, which could be different from that of the original data. Wethen propose a principled framework, termed Data Augmentation Optimized for GAN(DAG), to enable the use of augmented data in GAN training to improve thelearning of the original distribution. We provide theoretical analysis to showthat using our proposed DAG aligns with the original GAN in minimizing theJensen-Shannon (JS) divergence between the original distribution and modeldistribution. Importantly, the proposed DAG effectively leverages the augmenteddata to improve the learning of discriminator and generator. We conductexperiments to apply DAG to different GAN models: unconditional GAN,conditional GAN, self-supervised GAN and CycleGAN using datasets of naturalimages and medical images. The results show that DAG achieves consistent andconsiderable improvements across these models. Furthermore, when DAG is used insome GAN models, the system establishes state-of-the-art Frechet InceptionDistance (FID) scores. Our code is available.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×