Multimodal Controller for Generative Models
Class-conditional generative models are crucial tools for data generation from user-specified class labels. A number of existing approaches for class-conditional generative models require nontrivial modifications of existing architectures, in order to model conditional information fed into the model. In this paper, we introduce a new method called multimodal controller to generate multimodal data without introducing additional model parameters. With the proposed technique, the model can be trained easily from non-conditional generative models by simply attaching controllers at each layer. Each controller grants label-specific model parameters. Thus the proposed method does not require additional model complexity. In the absence of the controllers, our model reduces to non-conditional generative models. Numerical experiments demonstrate the effectiveness of our proposed method in comparison with those of the existing non-conditional and conditional generative models. Additionally, our numerical results demonstrate that a small portion (10 label-specific model parameters is required to generate class-conditional MNIST and FashionMNIST images.
READ FULL TEXT