SCGAN: Sampling and Clustering-Based Neural Architecture Search for GANs

Abstract

The evolutionary neural architecture search for generative adversarial networks (GANs) has demonstrated promising performance for generating high-quality images. However, two challenges persist, including the long search times and unstable search results. To alleviate these problems, this paper proposes a sampling and clustering-based neural architecture search algorithm for GANs, named SCGAN, which can significantly improve searching efficiency and enhance generation quality. Two improved strategies are proposed in SCGAN. First, a constraint sampling strategy is designed to limit the parameter capacity of architectures, which calculates their architecture size and discards those exceeding a reasonable parameter threshold. Second, a clustering selection strategy is applied in each architecture iteration, which integrates a decomposition selection mechanism and a hierarchical clustering mechanism to further improve search stability. Extensive experiments on the CIFAR-10 and STL-10 datasets demonstrated that SCGAN only requires 0.4 GPU days to find a promising GAN architecture in a vast search space including approximately 10$^{15}$ networks. Our best-found GAN outperformed those obtained by other neural architecture search methods with performance metric results (IS = 9.68$\pm$ 0.06, FID = 5.54) on CIFAR-10 and (IS = 12.12$\pm$ 0.13, FID = 12.54) on STL-10.