Semi-supervised GAN with similarity constraint for mode diversity

Mode collapse is a very common issue in Generative Adversarial Networks. To alleviate the mode collapse, we introduce a novel semi-supervised GAN-based generative model and propose a quantitative criterion to describe the degree of mode collapse. We design several schemes in the experiments to observe the effect of semi-supervised learning on mode collapse. In addition, the semi-supervised model can capture both supervised and unsupervised disentangled representation at the same time by introducing similarity constraint loss, so that generated image is higher-quality and more varied. The architecture leverages a few labels to control some factors on the class-conditional representation and captures other interpretable unsupervised representations with a large amount of unlabeled data. Both quantitative and visual results on the CIFAR-10 and SVHN datasets verify the ability of the proposed architecture.