迭代化代价函数及超参数可变的生成对抗网络  

Iterative cost function and variable parameter generative adversarial networks

作　　者：陈耀 宋晓宁[1] 於东军[2] Chen Yao;Song Xiaoning;Yu Dongjun(School of IoT Engineering,Jiangnan University,Wuxi 214122,China;School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122 [2]南京理工大学计算机科学与工程学院,江苏南京210094

出　　处：《南京理工大学学报》2019年第1期35-40,共6页Journal of Nanjing University of Science and Technology

基　　金：国家自然科学基金(61876072);国家重点研发计划子课题(2017YFC1601800);中国博士后科学基金特助(2018T110441);江苏省自然科学基金(BK20161135);江苏省"六大人才高峰"资助(XYDXX-012)

摘　　要：为了解决生成对抗网络(Generative adversarial networks,GAN)的训练难问题,该文在Wasserstein GAN(WGAN)方法基础上提出了迭代化代价函数及超参数可变的生成对抗网络。为了对原始WGAN中的惩罚项进行改进,用迭代的方法增加惩罚项代替原始随机选取的方法。针对WGAN中固定代价函数惩罚项的超参数,提出变动超参数策略,其变动的依据是仿分布和真实分布之间的距离。在MNIST手写字体数据集和CELEBA人脸数据集上的实验表明,与传统WGAN方法相比,该文方法在生成器的拟合速度上有了显著提高,充分验证了方法的有效性。In order to solve the difficult training problem of generative adversarial networks,this paper proposes an iterative cost function and variable parameter generative adversarial networks based on the Wasserstein GAN(WGAN)method. For the improvement of penalty items in the original WGAN,iterative methods are used to increase penalty instead of the original randomly selected method. Aiming at the hyper-parameter of penalty item of fixed cost function in WGAN,the strategy of changing hyper-parameter is put forward. The change is based on the distance between imitation distribution and real distribution. Experiments conducted on MNIST handwritten font datasets and CELEBA face datasets show the effectiveness of the proposed method as compared with the traditional WGAN,significantly improving the convergence speed of the generator.

关 键 词：生成对抗网络 迭代化代价函数 超参数可变 分布距离 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]