Depth and Complexity of Deep Generative Adversarial Networks

Abstract

Although generative adversarial networks (GANs) have achieved state-of-the-art results in generating realistic looking images, models often consist of neural networks with few layers compared to those for classification. We evaluate different architectures for GANs with varying depths using residual blocks with shortcut connections in order to train GANs with higher capacity. While training tend to oscillate and not benefit from additional capacity of naively stacked layers, we show that GANs are capable of generating images of higher visual fidelity with proper regularization and simple techniques such as minibatch discrimination. In particular, we show that an architecture similar to the standard GAN with residual blocks in the hidden layers consistently achieve higher inception scores than the standard model without noticeable mode collapse. The source code is made available on https://github.com/hvy/gan-complexity.