迁移学习+强化学习+生成对抗网络_迁移学习_强化学习

<html xmlns="http://www.w3.org/1999/xhtml"> <head> <meta charset="utf-8"> <meta name="generator" content="pdf2htmlEX"> <meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1"> <link rel="stylesheet" href="https://static.pudn.com/base/css/base.min.css"> <link rel="stylesheet" href="https://static.pudn.com/base/css/fancy.min.css"> <link rel="stylesheet" href="https://static.pudn.com/prod/directory_preview_static/62bf00ee9f2d63494a8104a0/raw.css"> <script src="https://static.pudn.com/base/js/compatibility.min.js"></script> <script src="https://static.pudn.com/base/js/pdf2htmlEX.min.js"></script> <script> try{ pdf2htmlEX.defaultViewer = new pdf2htmlEX.Viewer({}); }catch(e){} </script> <title></title> </head> <body> <div id="sidebar" style="display: none"> <div id="outline"> </div> </div> <div id="pf1" class="pf w0 h0" data-page-no="1"><div class="pc pc1 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/62bf00ee9f2d63494a8104a0/bg1.jpg"><div class="c x0 y1 w2 h2"><div class="t m0 x1 h3 y2 ff1 fs0 fc0 sc0 ls0 ws0">DCGAN</div><div class="t m0 x2 h4 y3 ff2 fs1 fc0 sc0 ls0 ws0">•</div><div class="t m0 x3 h5 y4 ff1 fs1 fc0 sc0 ls0 ws0">(1)<span class="_ _0"> </span><span class="ff3">生成<span class="_ _0"> </span></span>mnist<span class="_ _0"> </span><span class="ff3">图片： <span class="_ _0"> </span></span>python </div><div class="t m0 x3 h6 y5 ff1 fs1 fc0 sc0 ls0 ws0">main.py --input_height 28 -<span class="_ _1"></span>-</div><div class="t m0 x3 h6 y6 ff1 fs1 fc0 sc0 ls0 ws0">input_width 28 --<span class="_ _2"></span>output_height </div><div class="t m0 x3 h6 y7 ff1 fs1 fc0 sc0 ls0 ws0">28 --<span class="_ _2"></span>output_width 28 --dat<span class="_ _1"></span>aset </div><div class="t m0 x3 h6 y8 ff1 fs1 fc0 sc0 ls0 ws0">mnist --train --e<span class="_ _1"></span>poch 10 --</div><div class="t m0 x3 h6 y9 ff1 fs1 fc0 sc0 ls0 ws0">input_fname_patter<span class="_ _3"></span>n "*.jpg"</div><div class="t m0 x2 h4 ya ff2 fs1 fc0 sc0 ls0 ws0">•</div><div class="t m0 x3 h5 yb ff1 fs1 fc0 sc0 ls0 ws0">(2) <span class="_ _0"> </span><span class="ff3">生成人脸图片<span class="_ _1"></span>：<span class="_ _0"> </span><span class="ff1">python </span></span></div><div class="t m0 x3 h6 yc ff1 fs1 fc0 sc0 ls0 ws0">main.py --input_height 96 -<span class="_ _1"></span>-</div><div class="t m0 x3 h6 yd ff1 fs1 fc0 sc0 ls0 ws0">input_width 96 --<span class="_ _2"></span>output_height </div><div class="t m0 x3 h6 ye ff1 fs1 fc0 sc0 ls0 ws0">48 --<span class="_ _2"></span>output_width 48 --dat<span class="_ _1"></span>aset </div><div class="t m0 x3 h6 yf ff1 fs1 fc0 sc0 ls0 ws0">faces --c<span class="_ _1"></span>r<span class="_ _1"></span>op --t<span class="_ _1"></span>rain --epoch 300 </div><div class="t m0 x3 h6 y10 ff1 fs1 fc0 sc0 ls0 ws0">--input_fname_patte<span class="_ _1"></span>r<span class="_ _1"></span>n "*.jpg"</div><div class="t m0 x2 h4 y11 ff2 fs1 fc0 sc0 ls0 ws0">•</div><div class="t m0 x3 h5 y12 ff3 fs1 fc0 sc0 ls0 ws0">（人脸图片来<span class="_ _1"></span>源： </div><div class="t m0 x3 h6 y13 ff1 fs1 fc0 sc0 ls0 ws0">http://mmlab.ie.cuhk.edu.hk/pr<span class="_ _3"></span>o</div><div class="t m0 x3 h5 y14 ff1 fs1 fc0 sc0 ls0 ws0">jects/Cele<span class="_ _1"></span>bA<span class="_ _1"></span>.<span class="_ _1"></span>html<span class="_ _4"> </span><span class="ff3">）</span></div></div></div><div class="pi" data-data='{"ctm":[0.999990,0.000000,0.000000,0.999990,0.000000,0.000000]}'></div></div> </body> </html>

<div id="pf2" class="pf w0 h0" data-page-no="2"><div class="pc pc2 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/62bf00ee9f2d63494a8104a0/bg2.jpg"><div class="c x0 y1 w2 h2"><div class="t m0 x4 h7 y15 ff2 fs2 fc0 sc0 ls0 ws0">•</div><div class="t m0 x5 h8 y16 ff1 fs2 fc0 sc0 ls0 ws0">Deep <span class="_ _2"></span>Con<span class="_ _2"></span>volu<span class="_ _2"></span>tion<span class="_ _2"></span>al Ge<span class="_ _2"></span>nerat<span class="_ _2"></span>ive </div><div class="t m0 x5 h9 y17 ff1 fs2 fc0 sc0 ls0 ws0">A<span class="_ _1"></span>dvers<span class="_ _2"></span>arial N<span class="_ _2"></span>etwor<span class="_ _2"></span>ks<span class="_ _5"> </span><span class="ff3">是生<span class="_ _2"></span>成对抗网<span class="_ _2"></span>络</span></div><div class="t m0 x5 h9 y18 ff3 fs2 fc0 sc0 ls0 ws0">（<span class="_ _6"> </span><span class="ff1">Generat<span class="_ _2"></span>ive Adversaria<span class="_ _2"></span>l Net<span class="_ _2"></span>works<span class="_ _7"> </span></span>）的</div><div class="t m0 x5 h9 y19 ff3 fs2 fc0 sc0 ls0 ws0">一种延<span class="_ _2"></span>伸，将<span class="_ _2"></span>卷积网<span class="_ _2"></span>络引入<span class="_ _2"></span>到生成<span class="_ _2"></span>式模型<span class="_ _2"></span>当中</div><div class="t m0 x5 h9 y1a ff3 fs2 fc0 sc0 ls0 ws0">来做无<span class="_ _2"></span>监督的<span class="_ _2"></span>训练，<span class="_ _2"></span>利用卷<span class="_ _2"></span>积网络<span class="_ _2"></span>强大的<span class="_ _2"></span>特征</div><div class="t m0 x5 h9 y1b ff3 fs2 fc0 sc0 ls0 ws0">提取能<span class="_ _2"></span>力来提<span class="_ _2"></span>高生成<span class="_ _2"></span>网络的<span class="_ _2"></span>学习效<span class="_ _2"></span>果。</div><div class="t m0 x4 h7 y1c ff2 fs2 fc0 sc0 ls0 ws0">•</div><div class="t m0 x5 h9 y1d ff3 fs2 fc0 sc1 ls0 ws0">右图中<span class="_ _2"></span>：</div><div class="t m0 x4 h9 y1e ff3 fs2 fc0 sc1 ls0 ws0"> <span class="_ _6"> </span>上<span class="_ _2"></span>图<span class="_ _6"> </span><span class="ff4 sc0">a<span class="_ _6"> </span></span>是生成<span class="_ _2"></span>模型<span class="_ _6"> </span><span class="ff4 sc0">G<span class="_ _6"> </span></span>：<span class="sc0"> 输<span class="_ _2"></span>入（<span class="_ _6"> </span><span class="ff1">1<span class="_ _2"></span>00 <span class="_ _6"> </span></span>维噪<span class="_ _2"></span>声向</span></div><div class="t m0 x4 h9 y1f ff3 fs2 fc0 sc0 ls0 ws0">量<span class="_ _6"> </span><span class="ff1">Z<span class="_ _6"> </span></span>）到第<span class="_ _2"></span>一个<span class="_ _2"></span>层：全<span class="_ _2"></span>连接 <span class="_ _6"> </span><span class="ff1">1<span class="_ _2"></span>00 —> <span class="_ _2"></span>1024<span class="_ _5"> </span></span>，</div><div class="t m0 x4 h9 y20 ff3 fs2 fc0 sc0 ls0 ws0">然后再<span class="_ _2"></span>把<span class="_ _6"> </span><span class="ff1">1024<span class="_ _5"> </span></span>的一维<span class="_ _2"></span>向量<span class="_ _6"> </span><span class="ff1">r<span class="_ _1"></span>eshap<span class="_ _2"></span>e<span class="_ _6"> </span><span class="ff3">成<span class="_ _6"> </span></span>1024<span class="_ _5"> </span><span class="ff3">个</span></span></div><div class="t m0 x4 h9 y21 ff3 fs2 fc0 sc0 ls0 ws0">通道的<span class="_ _6"> </span><span class="ff1">4<span class="_ _2"></span>*4<span class="_ _6"> </span></span>的<span class="_ _6"> </span><span class="ff1">featu<span class="_ _2"></span>r<span class="_ _1"></span>e map<span class="_ _6"> </span><span class="ff3">。<span class="_ _2"></span>基本规<span class="_ _2"></span>律是生<span class="_ _2"></span>成</span></span></div><div class="t m0 x4 h9 y22 ff3 fs2 fc0 sc0 ls0 ws0">网络的<span class="_ _2"></span>每一个<span class="_ _2"></span>下一<span class="_ _2"></span>层是反卷<span class="_ _2"></span>积层<span class="_ _2"></span>，通道数<span class="_ _2"></span>减半<span class="_ _2"></span>，</div><div class="t m0 x4 h9 y23 ff3 fs2 fc0 sc0 ls0 ws0">图像尺<span class="_ _2"></span>寸加倍<span class="_ _2"></span>。</div><div class="t m0 x4 h9 y24 ff3 fs2 fc0 sc1 ls0 ws0"> <span class="_ _2"></span> <span class="_ _6"> </span>下图<span class="_ _6"> </span><span class="ff4 sc0">b<span class="_ _6"> </span></span>是判别<span class="_ _2"></span>模型<span class="_ _6"> </span><span class="ff4 sc0">D<span class="_ _6"> </span></span>：<span class="sc0">就<span class="_ _2"></span>是一<span class="_ _2"></span>个没有</span></div><div class="t m0 x4 h9 y25 ff1 fs2 fc0 sc0 ls0 ws0">poo<span class="_ _2"></span>lin<span class="_ _2"></span>g<span class="_ _6"> </span><span class="ff3">的全卷<span class="_ _2"></span>积网<span class="_ _2"></span>络，输<span class="_ _2"></span>出是一<span class="_ _2"></span>个标量<span class="_ _2"></span>，表示</span></div><div class="t m0 x4 h9 y26 ff3 fs2 fc0 sc0 ls0 ws0">输入数<span class="_ _2"></span>据属于<span class="_ _2"></span>训练<span class="_ _2"></span>数据而非<span class="_ _2"></span>生成<span class="_ _2"></span>样本的概<span class="_ _2"></span>率。</div></div></div><div class="pi" data-data='{"ctm":[0.999990,0.000000,0.000000,0.999990,0.000000,0.000000]}'></div></div>

<div id="pf3" class="pf w0 h0" data-page-no="3"><div class="pc pc3 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/62bf00ee9f2d63494a8104a0/bg3.jpg"><div class="c x0 y1 w2 h2"><div class="t m0 x4 ha y27 ff3 fs3 fc0 sc0 ls0 ws0">执行上述命令后会在<span class="_ _2"></span>本地目录下的<span class="_ _7"> </span><span class="ff1">samples<span class="_ _7"> </span></span>文件夹中生<span class="_ _2"></span>成一系列图片，如下所<span class="_ _2"></span>示：</div></div></div><div class="pi" data-data='{"ctm":[0.999990,0.000000,0.000000,0.999990,0.000000,0.000000]}'></div></div>