CAGAN : Constrained neural architecture search for GANs

Recently, a number of Neural Architecture Search (NAS) methods have been proposed to automate the design of Generative Adversarial Networks (GANs). However, due to the unstable training of GANs and the multi-model forgetting of one-shot NAS, the stability of embedding NAS into GANs is still not satisfactory. Thus, we propose a constrained evolutionary NAS method for GANs (called CAGAN) and design a first benchmark (NAS-GAN-Bench-101) for NAS in GANs. First, we constrain the sampling architecture size to steer the evolutionary search towards more promising and lightweight architectures. Subsequently, we propose a shape-constrained sampling strategy to select more reasonable architectures. Moreover, we present a multi-objective decomposition selection strategy to simultaneously consider the model shape, Inception Score (IS), and Fréchet Inception Distance (FID), which produces diverse superior generator candidates. CAGAN has been applied to unconditioned image generation tasks, in which the evolutionary search of GANs on the CIFAR-10 is completed in 0.35 GPU days. Our searched GANs showed promising results on the CIFAR-10 with (IS = 8.96 ± 0.06, FID = 9.45) and surpassed previous NAS-designed GANs on the STL-10 with (IS = 10.39 ± 0.13, FID = 19.34). Our source codes are available at https://github.com/fly2tortoise/CAGAN. © 2024 Elsevier B.V.