Training of Generative Adversarial Networks with Hybrid Evolutionary Optimization Technique

This paper proposes a new method for training Generative Adversarial Networks(GANs), named as Hybrid Evolutionary Optimization. The proposed method trains GAN networks by evolving generator set to reduce Fr ' echet Inception Distance. To overcome the problems such as non-convergence and mode collapse which are associated with conventional GANs, this paper uses an evolutionary algorithm to train in the initial iterations to stabilize the weights and followed by conventional optimization for remaining iterations. The efficiency of proposed method is evaluated with MNIST and celebA dataset. The results show that the proposed training algorithm converges faster than conventionally trained GANs. Moreover, the training with CPU and GPU is compared and analyzed