基于一种条件熵距离惩罚的生成式对抗网络  

Generative Adversarial Networks Based on Penalty of Conditional Entropy Distance

作　　者：谭宏卫 王国栋[1] 周林勇 张自力[1,3] TAN Hong-Wei;WANG Guo-Dong;ZHOU Lin-Yong;ZHANG Zi-Li(College of Computer and Information Science,Southwest University,Chongqing 400715,China;School of Mathematics and Statistics,Guizhou University of Finance and Economics,Guiyang 550025,China;School of Information Technology,Deakin University,Locked Bag 20000,Geelong,VIC 3220,Australia)

机构地区：[1]西南大学计算机与信息科学学院,重庆400715 [2]贵州财经大学数统学院,贵州贵阳550025 [3]School of Information Technology,Deakin University,Locked Bag 20000,Geelong,VIC 3220,Australia

出　　处：《软件学报》2021年第4期1116-1128,共13页Journal of Software

基　　金：国家自然科学基金(61732019)。

摘　　要：生成高质量的样本一直是生成式对抗网络(generative adversarial networks,简称GANs)领域的主要挑战之一.鉴于此,利用条件熵构建一种距离,并将此直接惩罚于GANs生成器目标函数,在尽可能保持熵不变的条件下,迫使生成分布逼近目标分布,从而大幅度地提高网络生成样本的质量.除此之外,还通过优化GANs的网络结构以及改变两个网络的初始化策略,以进一步提高GANs的训练效率.在多个数据集上的实验结果显示,所提出的算法显著提高了GANs生成样本的质量;尤其是在CIFAR10、STL10和Celeb A数据集上,将最佳的FID值从20.70、16.15、4.65分别降低到14.02、12.83、3.22.Generating high-quality samples is always one of the main challenges in generative adversarial networks(GANs) field. To this end, in this study, a GANs penalty algorithm is proposed, which leverages a constructed conditional entropy distance to penalize its generator. Under the condition of keeping the entropy invariant, the algorithm makes the generated distribution as close to the target distribution as possible and greatly improves the quality of the generated samples. In addition, to improve the training efficiency of GANs, the network structure of GANs is optimized and the initialization strategy of the two networks is changed. The experimental results on several datasets show that the penalty algorithm significantly improves the quality of generated samples. Especially, on the CIFAR10, STL10, and Celeb A datasets, the best FID value is reduced from 16.19, 14.10, 4.65 to 14.02, 12.83, and 3.22, respectively.

关 键 词：生成式对抗网络 条件熵距离 网络结构 样本多样性 图像生成 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]