Improved efficiency of sampling real data distribution, from the discriminator of a Generative Adversarial Network (GAN), using Metropolis-Hastings algorithm, and showed applications of proposed approach in reverse-KL divergence based GAN evaluation.