Conditional GANs with Auxiliary Discriminative Classifier

Conditional generative models aim to learn the underlying joint distribution of data and labels, and thus realize conditional generation. Among them, auxiliary classifier generative adversarial networks (AC-GAN) have been widely used, but suffer from the problem of low intra-class diversity on generated samples. In this paper, we point out that the fundamental reason is that the classifier of AC-GAN is generator-agnostic, and therefore cannot provide informative guidance to the generator to approximate the target distribution, resulting in minimization of conditional entropy that decreases the intra-class diversity. Motivated by this observation, we propose a novel conditional GAN with auxiliary \textit{discriminative} classifier (ADC-GAN) to resolve the problem of AC-GAN. Specifically, the proposed auxiliary \textit{discriminative} classifier becomes generator-aware by recognizing the labels of the real data and the generated data \textit{discriminatively}. Our theoretical analysis reveals that the generator can faithfully replicate the target distribution even without the original discriminator, making the proposed ADC-GAN robust to the hyper-parameter and stable on the training process. Extensive experimental results on synthetic and real-world datasets demonstrate the superiority of ADC-GAN on conditional generative modeling compared with competing methods.