Research And Implementation Of Generation Adversarial Networks Based On Multi-Generator

Since the proposal of generative adversarial networks,GAN has become a research hotspot of artificial intelligence.Various ideas have been proposed to improve the original GAN.GAN has outstanding effects on the field of computer image and vision,but in the meantime,it also has many shortcomings that haven't been solved yet,one of them who catch most attention is the problem of mode collapse.Broadly speaking,there are two schools of thought to address this issue:(1)improving the learning of GANs to reach better optima;and(2)explicitly enforcing GANs to capture diverse modes.Here we focus on the latter.We make some improvement to a GAN model which is based on multi-generator.The original model we selected has the problem of mode collapse in a single generator,besides the train phase of the selected model is very unstable,we propose three ideas to solve these problems and further improve the performance of the selected multi-generator model:1)Wasserstein distance is introduced to measure the difference between two data distributions,this can improve the stability of the model's train phase.Besides,Wasserstein distance can be used to measure the progress of the train phase.Wasserstein distance also has better performance used as loss function when dealing with mode collapse than the JSD divergence and KL divergence used in the selected model,thus this modification also can ease the problem of mode collapse in a single generator.2)We introduced residual block to build our neural network,this can reduce the number of parameters in the selected model,and improve the performance of the selected model.3)We modify the parameter sharing scheme used in the selected model.We unbind the parameters of the last level of networks in generators.Different categories of pictures often have different high-dimensional features,this modification let each generator handle features more independently.We also conduct extensive experiments to evaluate our improved model.We conducted experiments on MNIST first to preliminarily appraise the feasibility of our idea.Then we conducted experiments on Cifar10 and CelebA to further explore our improved model's performance on natural image databases,we introduce IS and FID Score for model evaluation.The experiments on these datasets demonstrate that our improved model can generate better samples,and can make the selected model more stable.