Training Generative Adversarial Networks in One Stage Chengchao Shen1, Youtan Yin1, Xinchao Wang2,5, Xubin Li3, Jie Song1,4,*, Mingli Song1 1Zhejiang University, 2National University of Singapore, 3Alibaba Group, 4Zhejiang Lab, 5Stevens Institute of Technology {chengchaoshen,youtanyin,sjie,brooksong}@zju.edu.cn,
[email protected],
[email protected] Abstract Two-Stage GANs (Vanilla GANs): Generative Adversarial Networks (GANs) have demon- Real strated unprecedented success in various image generation Fake tasks. The encouraging results, however, come at the price of a cumbersome training process, during which the gen- Stage1: Fix , Update 풢 풟 풢 풟 풢 풟 풢 풟 erator and discriminator풢 are alternately풟 updated in two 풢 풟 stages. In this paper, we investigate a general training Real scheme that enables training GANs efficiently in only one Fake stage. Based on the adversarial losses of the generator and discriminator, we categorize GANs into two classes, Sym- Stage 2: Fix , Update 풢 풟 metric GANs and풢 Asymmetric GANs,풟 and introduce a novel One-Stage GANs:풢 풟 풟 풢 풟 풢 풟 풢 gradient decomposition method to unify the two, allowing us to train both classes in one stage and hence alleviate Real the training effort. We also computationally analyze the ef- Fake ficiency of the proposed method, and empirically demon- strate that, the proposed method yields a solid 1.5× accel- Update and simultaneously eration across various datasets and network architectures. Figure 1: Comparison풢 of the conventional풟 Two-Stage GAN 풢 풟 Furthermore,풢 we show that the proposed풟 method is readily 풢 풟 풢 풟 풢 풟 training scheme (TSGANs) and the proposed One-Stage applicable to other adversarial-training scenarios, such as strategy (OSGANs).