Improving GAN Training with Probability Ratio Clipping and Sample Reweighting

Despite success on a wide range of problems related to vision, generative\nadversarial networks (GANs) often suffer from inferior performance due to\nunstable training, especially for text generation. To solve this issue, we\npropose a new variational GAN training framework which enjoys superior training\nstability. Our approach is inspired by a connection of GANs and reinforcement\nlearning under a variational perspective. The connection leads to (1)\nprobability ratio clipping that regularizes generator training to prevent\nexcessively large updates, and (2) a sample re-weighting mechanism that\nimproves discriminator training by downplaying bad-quality fake samples.\nMoreover, our variational GAN framework can provably overcome the training\nissue in many GANs that an optimal discriminator cannot provide any informative\ngradient to training generator. By plugging the training approach in diverse\nstate-of-the-art GAN architectures, we obtain significantly improved\nperformance over a range of tasks, including text generation, text style\ntransfer, and image generation.\n