1 2 3<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN"> 4<html xml:lang="en" xmlns="http://www.w3.org/1999/xhtml"> 5<head><TITLE>CDE-GAN(Shiming Chen)</TITLE> 6<META content="text/html; charset=UTF-8" http-equiv=Content-Type> 7 8<META content=IE=Edge,chrome=1 http-equiv=X-UA-Compatible> 9<LINK rel=stylesheet href="css/style.css" media=screen> 10<STYLE type=text/css> 11@media Print 12{ 13#STTBimg { 14 DISPLAY: none 15} 16#STTBimg2 { 17 DISPLAY: none 18} 19 } 20</STYLE> 21<style type="text/css"> 22 body{background:#FFFFFF;text-align:center;} 23 div{width:970px;margin:0 auto;background:#fff;text-align:left;} 24 25 26.STYLE2 {font-size: 18px} 27.STYLE6 {font-size: 24pt} 28.STYLE8 {color: #CC0000} 29.STYLE14 {color: #000000; font-size: 17px; font-weight: bold; } 30.STYLE15 {color: #000000} 31.STYLE16 {font-size: 22px} 32.STYLE20 {color: #FF0000; font-weight: bold; } 33.STYLE22 {color: #0000FF} 34</style> 35 36 37<META name=GENERATOR content="MSHTML 9.00.8112.16441"></head> 38<body class=" hasGoogleVoiceExt" screen_capture_injected="true"> 39 40<div class="container"> 41 <br><br> 42 <h1 class="STYLE6" style="TEXT-ALIGN: center">CDE-GAN: Cooperative Dual Evolution Based Generative Adversarial Network</H1> 43 <br> 44 <span class="STYLE2"><br> 45 </span> 46 <p style='text-align: center;'> 47 <span style="font-size: 18px"><a class='urllink' href='https://shiming-chen.github.io/' title='' rel='nofollow'><font color="#0000FF"><strong>Shiming Chen</strong></font></a><sup>1</sup>, 48 <a class='urllink' href='' title='' rel='nofollow'><font color="#0000FF"><strong>Wenjie Wang</strong></font></a><sup>1</sup>, 49 <a class='urllink' href='' title='' rel='nofollow'><font color="#0000FF"><strong>Beihao Xia</strong></font></a><sup>1</sup>, 50 <a class='urllink' href='http://bmal.hust.edu.cn/info/1005/1091.htm' title='' rel='nofollow'><font color="#0000FF"><strong>Xinge You</strong></font></a><sup>1</sup>, 51 <a class='urllink' href='http://bmal.hust.edu.cn/info/1005/1092.htm' title='' rel='nofollow'><font color="#0000FF"><strong>Qinmu Peng</strong></font></a><sup>1</sup>, 52 <a class='urllink' href='https://czh513.github.io/' title='' rel='nofollow'><font color="#0000FF"><strong>Zehong Cao</strong></font></a><sup>2</sup>, 53 and <a class='urllink' href='https://www.researchgate.net/profile/Weiping_Ding2' title='' rel='nofollow'><font color="#0000FF"><strong>Weiping Ding</strong></font></a><sup>3</sup> 54 </span><br> 55 56 </p> 57 <p class='vspace STYLE2' style='text-align: center;'> 58 <sup>1</sup>Huazhong University of Science and Technology (HUST), China </p> 59 <p class='vspace STYLE2' style='text-align: center;'> 60 <sup>2</sup>University of Tasmania (UTAS), Australia     <sup>3</sup>Nantong University (NTU), China </p> 61 <p class='vspace STYLE2' style='text-align: center;'> 62 <font color="#0000FF" font-style='italic' font-family='NimbusMonL-Regu'>{shimingchen, wangwj54, xbh_hust, youxg, pengqinmu}@hust.edu.cn   [email protected]   [email protected] </font> </p> 63 64 <div class=vspace> </div> 65 <div class=line> </div> 66 <br> 67 68 <h2 align="center">Abstract</h2> 69 <P align="justify">Generative adversarial networks (GANs) have been a popular deep generative model for real-world applications. 70 Despite many recent efforts on GANs that have been contributed, mode collapse and instability of GANs are still open problems caused by their adversarial 71 optimization difficulties. In this paper, motivated by the cooperative co-evolutionary algorithm, we propose a 72 \textit{Cooperative Dual Evolution based Generative Adversarial Network} (CDE-GAN) to circumvent these drawbacks. In essence, 73 CDE-GAN incorporates dual evolution with respect to the generator(s) and discriminators into a unified evolutionary adversarial framework to conduct 74 effective adversarial multi-objective optimization. Thus it exploits the complementary properties and injects dual mutation diversity into 75 training to steadily diversify the estimated density in capturing multi-modes and improve generative performance. Specifically, CDE-GAN decomposes the complex 76 adversarial optimization problem into two subproblems (generation and discrimination), and each subproblem is solved with a separated subpopulation 77 (\textit{E-Generators} and \textit{E-Discriminators}), evolved by its own evolutionary algorithm. Additionally, we further propose a \textit{Soft Mechanism} 78 to balance the trade-off between E-Generators and E-Discriminators to conduct steady training for CDE-GAN. Extensive experiments on one synthetic dataset and 79 three real-world benchmark image datasets demonstrate that the proposed C
79DE-GAN achieves a competitive and superior performance in generating good quality 80 and diverse samples over baselines. The code and more generated results are available at our project homepage 81 <a class='urllink' href=' https://shiming-chen.github.io/CDE-GAN-website/CDE-GAN.html' title='' 82 rel='nofollow'><font color="#0000FF"><strong> https://shiming-chen.github.io/CDE-GAN-website/CDE-GAN.html</strong></font></a>. </P> 83 <div class=line> </div> 84 <p></p> 85 <h2>Model pipeline </h2> 86 87 <div class="div-table-col-12" align='center'> 88 <center><img src="figures/CDEGAN-pipeline.jpg" width="950" height="350"></center> 89 <p></p> 90 <P align="justify"><strong>Figure:</strong>The pipeline of CDE-GAN. In brief, CDE-GAN decomposes the complex adversarial optimization problem into two 91 subproblems (generation and discrimination), and each subproblem is solved with a separated subpopulation (i.e., E-Generatorsand E-Discriminators), 92 evolved by an individual evolutionary algorithm (including individual Variations, Evaluations and Selections). The best offspring of E-Generators 93 and E-Discriminator are served as new parents to produce the next generation's individuals (i.e., children). Furthermore, a Soft Mechanism is 94 proposed to cooperate E-Generators and E-Discriminators to conduct effective adversarial training.</P> 95 96 97 </div> 98 <div class=line> </div> 99 <p></p> 100 <h2>Material </h2> 101 <form id="exp_texture"> 102 <div class="div-table" align='center'> 103 <div class="div-table-row"> 104 <div class="div-table-col-3" align='center' id="img-large" > 105 <center><a class=urllink title="" href="figures/CDE-GAN.pdf" rel=nofollow><img src="figures/CDE-GAN-page1.png" width="150" height="150" 106 class="layered-paper-big"></br></br></br><h3><font color="#0000FF"><strong>Paper</strong></font></h3></a></center> 107 </div> 108 <div class="div-table-col-3" align='center' id="img-large1" style="margin-left:150px"> 109 <center><a class=urllink title="" href="https://github.com/shiming-chen/CDE-GAN" rel=nofollow><img src="figures/github-figure.jpg" 110 width="150" height="150"></br></br></br><h3><font color="#0000FF"><strong>Code</strong><br></font></h3></a></center> 111 </div> 112 </div> 113 </div> 114 </form> 115 116 <div class=vspace></div> 117 118 <div> 119 <p>The supplementary can be downloaded <a class=urllink title="" href="figures/supplement.pdf" rel=nofollow><font color="#0000FF"><strong>here</strong></font></a> </p> 120 <p>If you wish to use our code, please cite the following paper : </p> 121 122 123 <div class=div-table-reference> 124 <strong>CDE-GAN: Cooperative Dual Evolution Based Generative Adversarial Network</strong><br> 125 Shiming Chen, Wenjie Wang, Beihao Xiao, Xinge You, Qinmu Peng, Zehong Cao, Weiping Ding<br> 126 <em>arXiv preprint arXiv: 2008.09388, </em> 2020<br> 127 </div> 128 </div> 129 130 <div class=vspace></div> 131 132 <div class=line> </div> 133 <p></p> 134 135 136 137 <h2>Evaluation</h2> 138 139 140 <p class="STYLE15">There are three different benchmark datasets used for evaluating our method, i.e., 141 <a class=urllink title="" href="http://www.cs.toronto.edu/~kriz/cifar.html" rel=nofollow><font size="4.2" font-weight="bold" color="#0000FF">CIFAR-10</font></a>, 142 <a class=urllink title="" href="https://github.com/fyu/lsun" rel=nofollow><font size="4.2" font-weight="bold" color="#0000FF">LSUN-Bedrooms</font></a>, 143 <!--<a class=urllink title="" href="" rel=nofollow><font size="4.2" font-weight="bold" color="#0000FF">Stacked-MNIST</font></a>,--> 144 and <a class=urllink title="" href="http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html " rel=nofollow><font size="4.2" color="#0000FF">CelebA</font></a>. 145 Some of these generated images are used in paper, in which supports better vision quality. Indeed, we also evaluate our method on synthesis dataset. </p> 146 147 148 <div class=line> </div> 149 150 <h3 class="STYLE16" id="exp1">Experiment 1: Experiments on CIFAR-10 for hyper-parameters analysis</h3> 151 <p class="STYLE15">Experiments on the CIFAR-10 dataset for hyper-parameters analysis. (a) Inception score evaluation for different CDE-GANs with various balance factor 152 γ={1,0.5,0.1,0.01}. (b) Inception score evaluation for different CDE-GANs with various numbers of discriminators I={1,2,4,8}.</p> 153 154 155 156 157 <form id="exp_texture"> 158 <div class="div-table" align='center'> 159 <div class="div-table-col-12" align='center'> 160 <center><img src="figures/experiment1.png" width="950" height="320"></center> 161 162 </div> 163 </div> 164 </form> 165 166 167 168 <div class=line> </div> 169 170 <h3 class="STYLE16">Experiment 2: Generative Performance Evaluation </h3> 171 <h4 class="STYLE16">2.1 Generated results on Synthetic dataset </h4> 172 <p class="STYLE15">Dynamic results of Gaussian kernel estimation over generator iteration for different GANs. For each pair of images, the left one is data distribution 173 (real data is represented in blue, generated data is represented in red), and the right one is KDE plots of the generated data corresponded to its left generated data. From 174 top to bottom, the rows are the results of original GAN, NS-GAN, LSGAN, E-GAN, and <strong>CDE-GAN (Ours)</strong>.</p> 175 176 <form id="exp_texture"> 177 <div class="div-table" align='center'> 178 <div class="div-table-col-12" align='center'> 179 <center><img src="figures/difgan-toy-mlp3.jpg" width="960" height="600"></center> 180 <center><h3>(a) Gaussian kernel estimation with MLP of 3 layers</h3></center> 181 182 183 </div> 184 <div class="div-table-col-12" align='center'> 185 <center><img src="figures/difgan-toy-mlp4.jpg" width="960" height="600"></center> 186 <center><h3>(b) Gaussian kernel estimation with MLP of 4 layers</h3></center> 187 </div> 188 189 190 </div> 191 </form> 192 193 <h4 class="STYLE16">2.2 Generated results on real-world datasets </h4> 194 195 <form id="exp_texture"> 196 <div class="div-table" align='center'> 197 <div class="div-table-col-12" align='center'> 198 <center><img src="figures/CelebA.png" width="960" height="960"></center> 199 <center><h3>(a) CelebA</h3></center> 200 </div> 201 <div class="div-table-col-12" align='center'> 202 <center><img src="figures/CIFAR.png" width="960" height="960"></center> 203 <center><h3>(b) CIFAR-10</h3></center> 204 </div> 205 <div class="div-table-col-12" align='center'> 206 <center><img src="figures/LSUN.png" width="960" height="960"></center> 207 <center><h3>(c) LSUN-Bedrooms</h3></center> 208 </div> 209 210 <!-- 211 <div class="div-table-col-12" align='center'> 212 <center><img src="figures/CIFAR-1.png" width="300" height="300"> 213     <img src="figures/CIFAR-2.png" width="300" height="300"> 214     <img src="figures/CIFAR-3.png" width="300" height="300"></center> 215 <center><h5>(a) CIFAR-10</h5></center> 216 </div> 217 218 <div class="div-table-col-12" align='center'> 219 <center><img src="figures/LSUN-1.png" width="300" height="300"> 220     <img src="figures/LSUN-2.png" width="300" height="300"> 221     <img src="figures/LSUN-3.png" width="300" height="300"></center> 222 <center><h5>(b) LSUN-Bedrooms</h5></center> 223 </div> 224 <div class="div-table-col-12" align='center'> 225 <center><img src="figures/CelebA-1.png" width="300" height="300"> 226     <img src="figures/CelebA-2.png" width="300" height="300"> 227     <img src="figures/CelebA-3.png" width="300" height="300"></center> 228 <center><h5>(c) CelebA</h5></center> 229 </div> 230 --> 231 232 </div> 233 </form> 234 235 236 <div class=line> </div> 237 <p> </p> 238 <p> </p> 239 <h3 class="STYLE16">Experiment 3: Comparisons with state-of-the-art methods </h3> 240 <h4 class="STYLE16">3.1 Quantitative comparison on CIFAR-10 dataset </h4> 241 <form id="exp_texture"> 242 <div class="div-table" align='center'> 243 <div class="div-table-col-12" align='center'> 244 <center><img src="figures/Table1.png" width="430" height="400">       <img src="figures/Table2.png" width="430" height="408"></center> 245 246 247 248 </div> 249 250 </div> 251 </form> 252 253 <h4 class="STYLE16">3.2 Qualitative comparison on CelebA </h4> 254 <form id="exp_texture"> 255 <p>Samples generated by different methods on various natural image datasets. The samples generated by different methods are provided by the original literatures, 256 i.e., MAD-GAN [23], Lipizzaner [24], Mustangs [27], Stabilizing-GAN [10], and acGAN [19]. 257 <div class="div-table" align='center'> 258 <div class="div-table-col-12" align='center'> 259 <center><img src="figures/Image-compared.png" width="900" height="600"> 260 261 </div> 262 263 </div> 264 </form> 265 266 267 268 269 270 271 272 273 <!--<center><img src="../../images/animation_hummingbird.gif" width=50% alt="" align="middle"></center> --> 274 275 <div class=line> </div> 276 <!-- <div class=line-seg> </div> --> 277 <p></p> 278 <h2>Acknowledgement</h2> 279 280 <p align="justify">This work was supported in part by the National Natural Science Foundation of China~(61571205 and 61772220), 281 the Key Program for International S\&T Cooperation Projects of China~(2016YFE0121200), the Special Projects for Technology Innovation of Hubei Province~(2018ACA135), 282 the Key Science and Technology Innovation Program of Hubei Province~(2017AAA017), the Natural Science Foundation of Hubei Province~(2018CFB691), 283 fund from Science, Technology and Innovation Commission of Shenzhen Municipality~(JCYJ20180305180637611, JCYJ20180305180804836 and JSGG20180507182030600). 284 We thank Dr.
285 <a class='urllink' href='https://www.sydney.edu.au/engineering/about/our-people/academic-staff/chaoyue-wang.html' title='' rel='nofollow'><font color="#0000FF"><strong> Chaoyue Wang</strong></font></a> 286 for his assistance with coding and theoretical suggestions.</p> 287 <div class=line> </div> 288 289 <p></p> 290 <h2>Reference</h2> 291 292 <p align="justify">[1] Chaoyue Wang et al. "Evolutionary generative adversarial networks." <em>IEEE Transactions on Evolutionary Computation (TEVC)</em>, 2019.</p> 293 <p align="justify">[2] Isabela Albuquerque et al. "Multi-objective training of Generative Adversarial Networks with multiple discriminators." <em>In ICML</em>, 2019.</p> 294 <p align="justify">[3] Jamal Toutouh et al. "Spatial evolutionary generative adversarial networks." <em>In GECCO</em>, 2019.</p> 295 <p align="justify">[4] Tom Schmiedlechner et al. "Towards distributed coevolutionary gans." <em>In AAAI</em>, 2018.</p> 296 <p align="justify">[5] Ishan Durugkar et al. "Generative multi-adversarial networks." <em>In ICLR</em>, 2017.</p> 297 <p align="justify">[6] Ian J. Goodfellow et al. "Generative adversarial nets." <em>In NIPS</em>, 2014.</p> 298 299 <div class=line> </div> 300 <h4> <center><a class='urllink' href='./CDE-GAN.html' title='' rel='nofollow'><font color="#0000FF">Top</font></a> </center></h4> 301</div> 302</body></html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.