Deep Automatic Control of Learning Rates for GANs

Toshiki Kamiya, Fumihiko Sakaue, Jun Sato

Abstract: In this paper, we propose a method for automatically controlling the learning rate of Generative Adversarial Networks (GANs) so as to stabilize the training of GANs. In recent years, GAN has been successful in various types of image generation tasks. Since GAN trains Generators and Discriminators adversarially, it is very important to keep the balance of their learning progress. However, it is known that the adjustment of learning rate of GAN is extremely difficult compared to conventional networks. Thus, we in this paper propose a method for predicting the future training progress of GANs from the current state of Generators and Discriminators, and for automatically controlling the learning rate of GANs appropriately. The proposed method has been tested using several different GANs, and the results show the proposed method can control the learning rate of GANs appropriately for a variety of tasks.