Parametric Adversarial Divergences are Good Losses for Generative Modeling