DEGAS: Differentiable Efficient Generator Search

TL;DR

DEGAS introduces a differentiable and efficient method for searching optimal GAN generators, outperforming previous RL-based approaches in quality and search speed, with broad applicability.

Contribution

This work presents DEGAS, a novel differentiable search strategy for GAN generators that improves efficiency and stability over prior reinforcement learning methods.

Findings

Outperforms RL-based GAN search methods in quality and speed.

Achieves higher inception scores on CIFAR-10 and STL datasets.

Provides a plug-and-play generator architecture for existing GAN frameworks.

Abstract

Network architecture search (NAS) achieves state-of-the-art results in various tasks such as classification and semantic segmentation. Recently, a reinforcement learning-based approach has been proposed for Generative Adversarial Networks (GANs) search. In this work, we propose an alternative strategy for GAN search by using a method called DEGAS (Differentiable Efficient GenerAtor Search), which focuses on efficiently finding the generator in the GAN. Our search algorithm is inspired by the differential architecture search strategy and the Global Latent Optimization (GLO) procedure. This leads to both an efficient and stable GAN search. After the generator architecture is found, it can be plugged into any existing framework for GAN training. For CTGAN, which we use in this work, the new model outperforms the original inception score results by 0.25 for CIFAR-10 and 0.77 for STL. It also gets better results than the RL based GAN search methods in shorter search time.