Improving Generalization and Stability of Generative Adversarial Networks