Regularizing Label-Augmented Generative Adversarial Networks Under Limited Data