Efficient Sampling for Generative Adversarial Networks with Coupling Markov Chains
Recently, sampling methods have been successfully applied to enhance the sample quality of Generative Adversarial Networks (GANs). However, in practice, they typically have poor sample efficiency because of the independent proposal sampling from the generator. In this work, we propose CMC-GAN, a novel sampling method with dependent proposals reparametrized through a Coupling Markov Chain (CMC) in the latent space of the generator. Theoretically, we prove that our CMC proposal can converge to the true data distribution, given a perfect discriminator. Empirically, extensive experiments on synthetic and real datasets demonstrate that our CMC-GAN largely improves the sample efficiency and obtains better sample quality simultaneously.
PDF Abstract