[2K4-ES-2] Machine learning: GAN

[2K4-ES-2-02] LCGAN: Conditional GAN with Multiple Discrete Classes

〇Sho Inoue¹, Tad Gonsalves¹ (1. Sophia University)

Keywords:GAN, VAE, Representation Learning, Discrete Variable , Latent Code

This paper introduces the way of generating data with some sets of classes by Latent Conditional Generative Adversarial Networks (LCGAN). LCGAN is conditional GAN which uses the latent code of Variational Autoencoder (VAE) as labels. The aim of this paper is generating the representation of continuous labels by not only continuous classes such as “age” but also discrete classes like “expressions” or “characteristics”. CelebA dataset which has also discrete annotation was used in this experiment. We could generate properly with 2 sets of classes by using the CelebA dataset. Further, since the LCGAN does not depend on the model structure, it can be easily extended to other GANs or VAEs.