Effective Lipschitz constraint enforcement for Wasserstein GAN training

Generative Adversarial Networks (GANs) are efficient frameworks for estimating generative model via adversarial process. However, GAN has known for suffering from training instability. Wasserstein GAN (WGAN) improves the training stability significantly but also brings an additional Lipschitz requirement for the critic network. To enforce the Lipschitz constraint, instead of weight clipping strategy, recent work adds a gradient penalty term to the critic loss. In this paper, we combine a more discriminative gradient penalty term with the importance weighting strategy and further propose more effective algorithms for Lipschitz constraint enforcement of the critic in WGAN. Our algorithms do not adding any computation burden.