Evaluating the distribution learning capabilities of GANs