On the Evaluation of Generative Adversarial Networks By Discriminative Models