 |
|
|
|
 |
 |
 |
|
 |
|
 |
|
|
谭宏卫,王国栋,周林勇,张自力.基于一种条件熵距离惩罚的生成式对抗网络研究.软件学报,0,(0):0 |
基于一种条件熵距离惩罚的生成式对抗网络研究 |
Research on Generative Adversarial Networks Based on a Penalty of a Conditional Entropy Distance |
投稿时间:2020-06-13 修订日期:2020-07-28 |
DOI:10.13328/j.cnki.jos.006156 |
中文关键词: 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 |
英文关键词:generative adversarial networks conditional entropy distance network structure sample diversity image generation |
基金项目:国家自然科学基金重点项目,面向需求不确定性的智能服务架构研究(61732019) |
作者 | 单位 | E-mail | 谭宏卫 | 西南大学 计算机与信息科学学院, 重庆 400715 贵州财经大学 数统学院, 贵州 贵阳 550025 | | 王国栋 | 西南大学 计算机与信息科学学院, 重庆 400715 | | 周林勇 | 贵州财经大学 数统学院, 贵州 贵阳 550025 | | 张自力 | 西南大学 计算机与信息科学学院, 重庆 400715 School of Information Technology, Deakin University, Locked Bag 20000, Geelong, VIC 3220, Australia | zhangzl@swu.edu.cn |
|
摘要点击次数: 375 |
全文下载次数: 237 |
中文摘要: |
生成高质量的样本一直是生成式对抗网络(Generative Adversarial Networks,GANs)领域的主要挑战之一.鉴于此,本文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,在尽可能地保持熵不变的条下,迫使生成分布逼近目标分布,从而大幅地提高网络生成样本的质量.除此之外,本文还通过优化GANs的网络结构以及改变两个网络的初始化策略,来进一步提高GANs的训练效率.在多个数据集上的实验结果显示,本文所提出的算法显著提高了GANs生成样本的质量;尤其在CIFAR10,STL10和CelebA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. |
英文摘要: |
Generating high-quality samples is always one of the most 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, we optimize the network structure of GANs and change the initialization strategy of the two networks. The experimental results on several datasets show that the penalty algorithm significantly improves the quality of generated samples. Especially, on the CIFAR10, STL10 and CelebA datasets, the best FID value was reduced from 16.19, 14.10, 4.65 to 14.02,12.83, 3.22, respectively. |
HTML 下载PDF全文 查看/发表评论 下载PDF阅读器 |
|
|
|
|
|
|
 |
|
|
|
|
 |
|
 |
|
 |
|