Optimizations of Ternary Generative Adversarial Networks

Generative adversarial networks (GANs), which can generate and transform data, have been attracting attention. However, the model must be lightweight and fast when applied in the field. In terms of model weight reduction, B-DCGAN, which restricts the value of the weights to {−1, +1}, has already been proposed. We propose a ternary GAN using the ternary representation {−1, 0, +1} and an $\alpha$ -layer which makes the learning between generator and discriminator more competitive. It succeeded in improving the quality of the output image considerably while maintaining the almost same memory usage as that of B-DCGAN.