Evaluating Implicit Generative Models With Large Samples