Stable and improved generative adversarial nets (GANS): A constructive survey
Guanghao Zhang, Enmei Tu, Dongshun Cui · 2017
In this paper, we present a general and applicable adversarial training framework based on a comprehensive survey, not limited to straightforward GANs related works, also including shallow neural networks and reinforcement learning. Concentrating on challenging face synthesis task, we summarize a stable training pipeline: 1) booting training procedure with noise injection; 2) fixing weights of fully connected layer in generator to improve performance further; 3) involving Markov decision module to dynamically choose learning rates of discriminator and generator respectively. Finally in experiments, we highlight a mutual evaluation criterion over entropy score based on a pre-trained classifier and manual voting.