Generative Counterfactual Introspection for Explainable Deep Learning