MetaGAN: An Adversarial Approach to Few-Shot Learning