Training Implicit Generative Models via an Invariant Statistical Loss