Enhance the Visual Representation via Discrete Adversarial Training