"Which Generative Adversarial Networks (GANs) generates the most plausible images?" has been a frequently asked question among researchers. To address this problem, we first propose an \emph{incomplete} U-statistics estimate of maximum mean discrepancy \textnormalMMDinc to measure the distribution discrepancy between generated and real images. \textnormalMMDinc enjoys the advantages of asymptotic normality, computation efficiency, and model agnosticity. We then propose a GANs analysis framework to select the "best" member in GANs family using the Post Selection Inference (PSI) with \textnormalMMDinc. In the experiments, we adopt the proposed framework on 7 GANs variants and compare their \textnormalMMDinc scores.
Live content is unavailable. Log in and register to view live content