PageSourceSearch

https://lyrichu.github.io/page/2/

html lyrichu.github.io collected 2026-10-03 09:32:49 UTC 366,045 bytes, 938 lines download raw bytes

1<!DOCTYPE html>
2<html>
3<head>
4  <meta charset="utf-8">
5  
6  <meta name="renderer" content="webkit">
7  <meta http-equiv="X-UA-Compatible" content="IE=edge" >
8  <link rel="dns-prefetch" href="http://Lyrichu.github.io">
9  <title>Lyrichu&#39;s Blog</title>
10  <meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=1">
11  <meta name="description" content="To be the best yourself!">
12<meta name="keywords" content="coding,math,machine learning">
13<meta property="og:type" content="website">
14<meta property="og:title" content="Lyrichu&#39;s Blog">
15<meta property="og:url" content="http://Lyrichu.github.io/page/2/index.html">
16<meta property="og:site_name" content="Lyrichu&#39;s Blog">
17<meta property="og:description" content="To be the best yourself!">
18<meta property="og:locale" content="zh-CN">
19<meta name="twitter:card" content="summary">
20<meta name="twitter:title" content="Lyrichu&#39;s Blog">
21<meta name="twitter:description" content="To be the best yourself!">
22  
23    <link rel="alternative" href="/atom.xml" title="Lyrichu&#39;s Blog" type="application/atom+xml">
24  
25  
26    <link rel="icon" href="/favicon.png">
27  
28  <link rel="stylesheet" type="text/css" href="/./main.0cf68a.css">
29  <style type="text/css">
30  
31    #container.show {
32      background: linear-gradient(200deg,#a0cfe4,#e8c37e);
33    }
34  </style>
35  
36
37  
38
39</head>
40
41<body>
42  <div id="container" q-class="show:isCtnShow">
43    <canvas id="anm-canvas" class="anm-canvas"></canvas>
44    <div class="left-col" q-class="show:isShow">
45      
46<div class="overlay" style="background: #4d4d4d"></div>
47<div class="intrude-less">
48	<header id="header" class="inner">
49		<a href="/" class="profilepic">
50			<img src="http://t1.aixinxi.net/o_1cegci9ej58d14rh1ts1siqcraa.jpg-w.jpg" class="js-avatar">
51		</a>
52		<hgroup>
53		  <h1 class="header-author"><a href="/">lyrichu</a></h1>
54		</hgroup>
55		
56		<p class="header-subtitle">Love life,love coding!</p>
57		
58
59		<nav class="header-menu">
60			<ul>
61			
62				<li><a href="/">主页</a></li>
63	        
64			</ul>
65		</nav>
66		<nav class="header-smart-menu">
67    		
68    			
69    			<a q-on="click: openSlider(e, 'innerArchive')" href="javascript:void(0)">所有文章</a>
70    			
71            
72    			
73    			<a q-on="click: openSlider(e, 'friends')" href="javascript:void(0)">友情链接</a>
74    			
75            
76    			
77    			<a q-on="click: openSlider(e, 'aboutme')" href="javascript:void(0)">关于我</a>
78    			
79            
80		</nav>
81		<nav class="header-nav">
82			<div class="social">
83				
84					<a class="github" target="_blank" href="http://www.github.com/Lyrichu" title="github"><i class="icon-github"></i></a>
85		        
86					<a class="weibo" target="_blank" href="#" title="weibo"><i class="icon-weibo"></i></a>
87		        
88					<a class="rss" target="_blank" href="#" title="rss"><i class="icon-rss"></i></a>
89		        
90					<a class="zhihu" target="_blank" href="https://www.zhihu.com/people/hu-cheng-chun-71" title="zhihu"><i class="icon-zhihu"></i></a>
91		        
92			</div>
93		</nav>
94	</header>		
95</div>
96
97    </div>
98    <div class="mid-col" q-class="show:isShow,hide:isShow|isFalse">
99      
100<nav id="mobile-nav">
101  	<div class="overlay js-overlay" style="background: #4d4d4d"></div>
102	<div class="btnctn js-mobile-btnctn">
103  		<div class="slider-trigger list" q-on="click: openSlider(e)"><i class="icon icon-sort"></i></div>
104	</div>
105	<div class="intrude-less">
106		<header id="header" class="inner">
107			<div class="profilepic">
108				<img src="http://t1.aixinxi.net/o_1cegci9ej58d14rh1ts1siqcraa.jpg-w.jpg" class="js-avatar">
109			</div>
110			<hgroup>
111			  <h1 class="header-author js-header-author">lyrichu</h1>
112			</hgroup>
113			
114			<p class="header-subtitle"><i class="icon icon-quo-left"></i>Love life,love coding!<i class="icon icon-quo-right"></i></p>
115			
116			
117			
118				
119			
120			
121			
122			<nav class="header-nav">
123				<div class="social">
124					
125						<a class="github" target="_blank" href="http://www.github.com/Lyrichu" title="github"><i class="icon-github"></i></a>
126			        
127						<a class="weibo" target="_blank" href="#" title="weibo"><i class="icon-weibo"></i></a>
128			        
129						<a class="rss" target="_blank" href="#" title="rss"><i class="icon-rss"></i></a>
130			        
131						<a class="zhihu" target="_blank" href="https://www.zhihu.com/people/hu-cheng-chun-71" title="zhihu"><i class="icon-zhihu"></i></a>
132			        
133				</div>
134			</nav>
135
136			<nav class="header-menu js-header-menu">
137				<ul style="width: 50%">
138				
139				
140					<li style="width: 100%"><a href="/">主页</a></li>
141		        
142				</ul>
143			</nav>
144		</header>				
145	</div>
146	<div class="mobile-mask" style="display:none" q-show="isShow"></div>
147</nav>
148
149      <div id="wrapper" class="body-wrap">
150        <div class="menu-l">
151          <div class="canvas-wrap">
152            <canvas data-colors="#eaeaea" data-sectionHeight="100" data-contentId="js-content" id="myCanvas1" class="anm-canvas"></canvas>
153          </div>
154          <div id="js-content" class="content-ll">
155            
156  
157    <article id="post-DCGAN代码简单解读" class="article article-type-post  article-index" itemscope itemprop="blogPost">
158  <div class="article-inner">
159    
160      <header class="article-header">
161        
162  
163    <h1 itemprop="name">
164      <a class="article-title" href="/2018/05/27/DCGAN代码简单解读/">DCGAN代码简单解读</a>
165    </h1>
166  
167
168        
169        <a href="/2018/05/27/DCGAN代码简单解读/" class="archive-article-date">
170  	<time datetime="2018-05-27T09:21:51.000Z" itemprop="datePublished"><i class="icon-calendar icon"></i>2018-05-27</time>
171</a>
172        
173      </header>
174    
175    <div class="article-entry" itemprop="articleBody">
176      
177        <h3 id="nbsp-nbsp-nbsp-nbsp-之前在DCGAN文章简单解读里说明了DCGAN的原理。本次来实现一个DCGAN-并在数据集上实际测试它的效果。本次的代码来自github开源代码DCGAN-tensorflow-感谢carpedm20的贡献"><a href="#nbsp-nbsp-nbsp-nbsp-之前在DCGAN文章简单解读里说明了DCGAN的原理。本次来实现一个DCGAN-并在数据集上实际测试它的效果。本次的代码来自github开源代码DCGAN-tensorflow-感谢carpedm20的贡献" class="headerlink" title="&nbsp;&nbsp;&nbsp;&nbsp;之前在DCGAN文章简单解读里说明了DCGAN的原理。本次来实现一个DCGAN,并在数据集上实际测试它的效果。本次的代码来自github开源代码DCGAN-tensorflow,感谢carpedm20的贡献!"></a>&nbsp;&nbsp;&nbsp;&nbsp;之前在<a href="http://www.movieb2b.com/2018/05/18/dcgan-%E6%96%87%E7%AB%A0%E4%BB%A5%E5%8F%8A%E4%BB%A3%E7%A0%81%E7%AE%80%E5%8D%95%E8%A7%A3%E8%AF%BB/" target="_blank" rel="noopener">DCGAN文章简单解读</a>里说明了DCGAN的原理。本次来实现一个DCGAN,并在数据集上实际测试它的效果。本次的代码来自github开源代码<a href="https://github.com/carpedm20/DCGAN-tensorflow" target="_blank" rel="noopener">DCGAN-tensorflow</a>,感谢<a href="https://github.com/carpedm20" target="_blank" rel="noopener">carpedm20</a>的贡献!</h3><h3 id="1-代码结构"><a href="#1-代码结构" class="headerlink" title="1. 代码结构"></a>1. 代码结构</h3><h3 id="nbsp-nbsp-nbsp-nbsp-代码结构如下图1所示"><a href="#nbsp-nbsp-nbsp-nbsp-代码结构如下图1所示" class="headerlink" title="&nbsp;&nbsp;&nbsp;&nbsp;代码结构如下图1所示:"></a>&nbsp;&nbsp;&nbsp;&nbsp;代码结构如下图1所示:</h3><p><div align="center"><br><img src="http://t1.aixinxi.net/o_1cebffbrgs8j1icd1d0p1blo6uca.png-w.jpg"><br></div></p>
178<center>图1 代码结构</center>
179
180<p>我们主要关注的文件为<strong>download.py</strong>,<strong>main.py</strong>,<strong>model.py</strong>,<strong>ops.py</strong>以及<strong>utils.py</strong>。其实看文件名字就大概可以猜出各个文件的作用了。</p>
181<ul>
182<li>
182download.py主要下载数据集到本地,这里我们需要下载三个数据集:<a href="http://yann.lecun.com/exdb/mnist/" target="_blank" rel="noopener">MNIST</a>,<a href="http://lsun.cs.princeton.edu/2016/" target="_blank" rel="noopener">lsun</a>以及<a href="http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html" target="_blank" rel="noopener">celebA</a>。</li>
183<li>main.py是主函数,用于配置命令行参数以及模型的训练和测试。</li>
184<li>model.py 是定义DCGAN模型的地方,也是我们要重点关注的代码。</li>
185<li>ops.py 定义了很多构造模型的重要函数,比如<strong>batch_norm</strong>(BN操作),<strong>conv2d</strong>(卷积操作),<strong>deconv2d</strong>(翻卷积操作)等。</li>
186<li><p>utils.py 定义很多有用的全局辅助函数。</p>
187<h3 id="2-代码简单解读"><a href="#2-代码简单解读" class="headerlink" title="2. 代码简单解读"></a>2. 代码简单解读</h3><h4 id="2-1-download-py"><a href="#2-1-download-py" class="headerlink" title="2.1 download.py"></a>2.1 <strong>download.py</strong></h4><h4 id="download-py代码如下"><a href="#download-py代码如下" class="headerlink" title="download.py代码如下:"></a>download.py代码如下:</h4><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br><span class="line">44</span><br><span class="line">45</span><br><span class="line">46</span><br><span class="line">47</span><br><span class="line">48</span><br><span class="line">49</span><br><span class="line">50</span><br><span class="line">51</span><br><span class="line">52</span><br><span class="line">
18753</span><br><span class="line">54</span><br><span class="line">55</span><br><span class="line">56</span><br><span class="line">57</span><br><span class="line">58</span><br><span class="line">59</span><br><span class="line">60</span><br><span class="line">61</span><br><span class="line">62</span><br><span class="line">63</span><br><span class="line">64</span><br><span class="line">65</span><br><span class="line">66</span><br><span class="line">67</span><br><span class="line">68</span><br><span class="line">69</span><br><span class="line">70</span><br><span class="line">71</span><br><span class="line">72</span><br><span class="line">73</span><br><span class="line">74</span><br><span class="line">75</span><br><span class="line">76</span><br><span class="line">77</span><br><span class="line">78</span><br><span class="line">79</span><br><span class="line">80</span><br><span class="line">81</span><br><span class="line">82</span><br><span class="line">83</span><br><span class="line">84</span><br><span class="line">85</span><br><span class="line">86</span><br><span class="line">87</span><br><span class="line">88</span><br><span class="line">89</span><br><span class="line">90</span><br><span class="line">91</span><br><span class="line">92</span><br><span class="line">93</span><br><span class="line">94</span><br><span class="line">95</span><br><span class="line">96</span><br><span class="line">97</span><br><span class="line">98</span><br><span class="line">99</span><br><span class="line">100</span><br><span class="line">101</span><br><span class="line">102</span><br><span class="line">103</span><br><span class="line">104</span><br><span class="line">105</span><br><span class="line">106</span><br><span class="line">107</span><br><span class="line">108</span><br><span class="line">109</span><br><span class="line">110</span><br><span class="line">111</span><br><span class="line">112</span><br><span class="line">113</span><br><span class="line">114</span><br><span class="line">115</span><br><span class="line">116</span><br><span class="line">117</span><br><span class="line">118</span><br><span class="line">119</span><br><span class="line">120</span><br><span class="line">121</span><br><span class="line">122</span><br><span class="line">123</span><br><span class="line">124</span><br><span class="line">125</span><br><span class="line">126</span><br><span class="line">127</span><br><span class="line">128</span><br><span class="line">129</span><br><span class="line">130</span><br><span class="line">131</span><br><span class="line">132</span><br><span class="line">133</span><br><span class="line">134</span><br><span class="line">135</span><br><span class="line">136</span><br><span class="line">137</span><br><span class="line">138</span><br><span class="line">139</span><br><span class="line">140</span><br><span class="line">141</span><br><span class="line">142</span><br><span class="line">143</span><br><span class="line">144</span><br><span class="line">145</span><br><span class="line">146</span><br><span class="line">147</span><br><span class="line">148</span><br><span class="line">149</span><br><span class="line">150</span><br><span class="line">151</span><br><span class="line">152</span><br><span class="line">153</span><br><span class="line">154</span><br><span class="line">155</span><br><span class="line">156</span><br><span class="line">157</span><br><span class="line">158</span><br><span class="line">159</span><br><span class="line">160</span><br><span class="line">161</span><br><span class="line">162</span><br><span class="line">163</span><br><span class="line">164</span><br><span class="line">165</span><br><span class="line">166</span><br><span class="line">167</span><br><span class="line">168</span><br><span class="line">169</span><br><span class="line">170</span><br><span class="line">171</span><br><span class="line">172</span><br><span class="line">173</span><br><span class="line">174</span><br><span class="line">175</span><br><span class="line">176</span><br><span class="line">177</span><br><span class="line">178</span><br><span class="line">179</span><br><span class="line">180</span><br><span class="line">181</span><br><span class="line">182</span><br><span class="line">183</span><br><span class="line">184</span><br><span class="line">185</span><br><span class="line">186</span><br><span class="line">187</span><br></pre></td><td class="code"><pre><span class="line"><span class="string">"""</span></span><br><span class="line"><span class="string">Modification of https://github.com/stanfordnlp/treelstm/blob/master/scripts/download.py</span></span><br><span class="line"><span class="string"></span></span><br><span class="line"><span class="string">Downloads the following:</span></span><br><span class="line"><span class="string">- Celeb-A dataset</span></span><br><span class="line"><span class="string">- LSUN dataset</span></span><br><span class="line"><span class="string">- MNIST dataset</span></span><br><span class="line"><span class="string">"""</span></span><br><span class="line"></span><br><span class="line"><span class="keyword">from</span> __future__ <span class="keyword">
187import</span> print_function</span><br><span class="line"><span class="keyword">import</span> os</span><br><span class="line"><span class="keyword">import</span> sys</span><br><span class="line"><span class="keyword">import</span> gzip</span><br><span class="line"><span class="keyword">import</span> json</span><br><span class="line"><span class="keyword">import</span> shutil</span><br><span class="line"><span class="keyword">import</span> zipfile</span><br><span class="line"><span class="keyword">import</span> argparse</span><br><span class="line"><span class="keyword">import</span> requests</span><br><span class="line"><span class="keyword">import</span> subprocess</span><br><span class="line"><span class="keyword">from</span> tqdm <span class="keyword">import</span> tqdm</span><br><span class="line"><span class="keyword">from</span> six.moves <span class="keyword">import</span> urllib</span><br><span class="line"></span><br><span class="line">parser = argparse.ArgumentParser(description=<span class="string">'Download dataset for DCGAN.'</span>)</span><br><span class="line">parser.add_argument(<span class="string">'datasets'</span>, metavar=<span class="string">'N'</span>, type=str, nargs=<span class="string">'+'</span>, choices=[<span class="string">'celebA'</span>, <span class="string">'lsun'</span>, <span class="string">'mnist'</span>],</span><br><span class="line">           help=<span class="string">'name of dataset to download [celebA, lsun, mnist]'</span>)</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">download</span><span class="params">(url, dirpath)</span>:</span></span><br><span class="line">  filename = url.split(<span class="string">'/'</span>)[<span class="number">-1</span>]</span><br><span class="line">  filepath = os.path.join(dirpath, filename)</span><br><span class="line">  u = urllib.request.urlopen(url)</span><br><span class="line">  f = open(filepath, <span class="string">'wb'</span>)</span><br><span class="line">  filesize = int(u.headers[<span class="string">"Content-Length"</span>])</span><br><span class="line">  print(<span class="string">"Downloading: %s Bytes: %s"</span> % (filename, filesize))</span><br><span class="line"></span><br><span class="line">  downloaded = <span class="number">0</span></span><br><span class="line">  block_sz = <span class="number">8192</span></span><br><span class="line">  status_width = <span class="number">70</span></span><br><span class="line">  <span class="keyword">while</span> <span class="keyword">True</span>:</span><br><span class="line">    buf = u.read(block_sz)</span><br><span class="line">    <span class="keyword">if</span> <span class="keyword">not</span> buf:</span><br><span class="line">      print(<span class="string">''</span>)</span><br><span class="line">      <span class="keyword">break</span></span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      print(<span class="string">''</span>, end=<span class="string">'\r'</span>)</span><br><span class="line">    downloaded += len(buf)</span><br><span class="line">    f.write(buf)</span><br><span class="line">    status = ((<span class="string">"[%-"</span> + str(status_width + <span class="number">1</span>) + <span class="string">"s] %3.2f%%"</span>) %</span><br><span class="line">      (<span class="string">'='</span> * int(float(downloaded) / filesize * status_width) + <span class="string">'&amp;gt;'</span>, downloaded * <span class="number">100.</span> / filesize))</span><br><span class="line">    print(status, end=<span class="string">''</span>)</span><br><span class="line">    sys.stdout.flush()</span><br><span class="line">  f.close()</span><br><span class="line">  <span class="keyword">return</span> filepath</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">download_file_from_google_drive</span><span class="params">(id, destination)</span>:</span></span><br><span class="line">  URL = <span class="string">"https://docs.google.com/uc?export=download"</span></span><br><span class="line">  session = requests.Session()</span><br><span class="line"></span><br><span class="line">  response = session.get(URL, params=&#123; <span class="string">
187'id'</span>: id &#125;, stream=<span class="keyword">True</span>)</span><br><span class="line">  token = get_confirm_token(response)</span><br><span class="line"></span><br><span class="line">  <span class="keyword">if</span> token:</span><br><span class="line">    params = &#123; <span class="string">'id'</span> : id, <span class="string">'confirm'</span> : token &#125;</span><br><span class="line">    response = session.get(URL, params=params, stream=<span class="keyword">True</span>)</span><br><span class="line"></span><br><span class="line">  save_response_content(response, destination)</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">get_confirm_token</span><span class="params">(response)</span>:</span></span><br><span class="line">  <span class="keyword">for</span> key, value <span class="keyword">in</span> response.cookies.items():</span><br><span class="line">    <span class="keyword">if</span> key.startswith(<span class="string">'download_warning'</span>):</span><br><span class="line">      <span class="keyword">return</span> value</span><br><span class="line">  <span class="keyword">return</span> <span class="keyword">None</span></span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">save_response_content</span><span class="params">(response, destination, chunk_size=<span class="number">32</span>*<span class="number">1024</span>)</span>:</span></span><br><span class="line">  total_size = int(response.headers.get(<span class="string">'content-length'</span>, <span class="number">0</span>))</span><br><span class="line">  <span class="keyword">with</span> open(destination, <span class="string">"wb"</span>) <span class="keyword">as</span> f:</span><br><span class="line">    <span class="comment"># 显示进度条</span></span><br><span class="line">    <span class="keyword">for</span> chunk <span class="keyword">in</span> tqdm(response.iter_content(chunk_size), total=total_size,</span><br><span class="line">              unit=<span class="string">'B'</span>, unit_scale=<span class="keyword">True</span>, desc=destination):</span><br><span class="line">      <span class="keyword">if</span> chunk: <span class="comment"># filter out keep-alive new chunks</span></span><br><span class="line">        f.write(chunk)</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">unzip</span><span class="params">(filepath)</span>:</span></span><br><span class="line">  print(<span class="string">"Extracting: "</span> + filepath)</span><br><span class="line">  dirpath = os.path.dirname(filepath)</span><br><span class="line">  <span class="keyword">with</span> zipfile.ZipFile(filepath) <span class="keyword">as</span> zf:</span><br><span class="line">    zf.extractall(dirpath)</span><br><span class="line">  os.remove(filepath)</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">download_celeb_a</span><span class="params">(dirpath)</span>:</span></span><br><span class="line">  data_dir = <span class="string">'celebA'</span></span><br><span class="line">  <span class="comment"># ./data/celebA</span></span><br><span class="line">  <span class="keyword">if</span> os.path.exists(os.path.join(dirpath, data_dir)):</span><br><span class="line">    print(<span class="string">'Found Celeb-A - skip'</span>)</span><br><span class="line">    <span class="keyword">return</span></span><br><span class="line"></span><br><span class="line">  filename, drive_id  = <span class="string">"img_align_celeba.zip"</span>, <span class="string">"0B7EVK8r0v71pZjFTYXZWM3FlRnM"</span></span><br><span class="line">  <span class="comment"># ./data/img_align_celeba.zip</span></span><br><span class="line">  save_path = os.path.join(dirpath, filename)</span><br><span class="line">  <span class="keyword">if</span> os.path.exists(save_path):</span><br><span class="line">    print(<span class="string">'[*] &#123;&#125; already exists'</span>.format(save_path)) <span class="comment"># 文件已经存在</span></span><br><span class="line">  <span class="keyword">else</span>:</span><br><span class="line">    download_file_from_google_drive(drive_id, save_path)</span><br><span class="line"></span><br><span class="line">  zip_dir = <span class="string">''</span></span><br><span class="line">  <span class="keyword">with</span> zipfile.ZipFile(save_path) <span class="keyword">as</span> zf:</span><br><span class="line">    zip_dir = zf.namelist()[<span class="number">0</span>] <span class="comment"># 解压以后默认文件夹的名字</span></span><br><span class="line">    zf.extractall(dirpath) <span class="comment"># 提取文件到该文件夹</span></span><br><span class="line">  os.remove(save_path) <span class="comment"># 移除压缩文件</span></span><br><span class="line">  <span class="comment"># 重命名文件夹</span></span><br><span class="line">  os.rename(os.path.join(dirpath, zip_dir), os.path.join(dirpath, data_dir))</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">_list_categories</span><span class="params">(tag)</span>:</span></span><br><span class="line">  url = <span class="string">'http://lsun.cs.princeton.edu/htbin/list.cgi?tag='</span> + tag</span><br><span class="line">  f = urllib.request.urlopen(url)</span><br><span class="line">  <span class="keyword">return</span> json.loads(f.read())</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">_download_lsun</span><span class="params">(out_dir, category, set_name, tag)</span>:</span></span><br><span class="line">  <span class="comment"># locals(),Return a dictionary containing the current scope's local variables</span></span><br><span class="line">  url = <span class="string">'http://lsun.cs.princeton.edu/htbin/download.cgi?tag=&#123;tag&#125;'</span> \</span><br><span class="line">
187      <span class="string">'&amp;amp;category=&#123;category&#125;&amp;amp;set=&#123;set_name&#125;'</span>.format(**locals())</span><br><span class="line">  print(url)</span><br><span class="line">  <span class="keyword">if</span> set_name == <span class="string">'test'</span>:</span><br><span class="line">    out_name = <span class="string">'test_lmdb.zip'</span></span><br><span class="line">  <span class="keyword">else</span>:</span><br><span class="line">    out_name = <span class="string">'&#123;category&#125;_&#123;set_name&#125;_lmdb.zip'</span>.format(**locals())</span><br><span class="line">  <span class="comment"># out_path:./data/lsun/xxx.zip</span></span><br><span class="line">  out_path = os.path.join(out_dir, out_name)</span><br><span class="line">  cmd = [<span class="string">'curl'</span>, url, <span class="string">'-o'</span>, out_path]</span><br><span class="line">  print(<span class="string">'Downloading'</span>, category, set_name, <span class="string">'set'</span>)</span><br><span class="line">  <span class="comment"># 调用linux命令</span></span><br><span class="line">  subprocess.call(cmd)</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">download_lsun</span><span class="params">(dirpath)</span>:</span></span><br><span class="line">  data_dir = os.path.join(dirpath, <span class="string">'lsun'</span>)</span><br><span class="line">  <span class="keyword">if</span> os.path.exists(data_dir):</span><br><span class="line">    print(<span class="string">'Found LSUN - skip'</span>)</span><br><span class="line">    <span class="keyword">return</span></span><br><span class="line">  <span class="keyword">else</span>:</span><br><span class="line">    os.mkdir(data_dir)</span><br><span class="line"></span><br><span class="line">  tag = <span class="string">'latest'</span></span><br><span class="line">  <span class="comment">#categories = _list_categories(tag)</span></span><br><span class="line">  categories = [<span class="string">'bedroom'</span>]</span><br><span class="line"></span><br><span class="line">  <span class="keyword">for</span> category <span class="keyword">in</span> categories:</span><br><span class="line">    _download_lsun(data_dir, category, <span class="string">'train'</span>, tag)</span><br><span class="line">    _download_lsun(data_dir, category, <span class="string">'val'</span>, tag)</span><br><span class="line">  _download_lsun(data_dir, <span class="string">''</span>, <span class="string">'test'</span>, tag)</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">download_mnist</span><span class="params">(dirpath)</span>:</span></span><br><span class="line">  data_dir = os.path.join(dirpath, <span class="string">'mnist'</span>)</span><br><span class="line">  <span class="keyword">if</span> os.path.exists(data_dir):</span><br><span class="line">    print(<span class="string">'Found MNIST - skip'</span>)</span><br><span class="line">    <span class="keyword">return</span></span><br><span class="line">  <span class="keyword">else</span>:</span><br><span class="line">    os.mkdir(data_dir)</span><br><span class="line">  url_base = <span class="string">'http://yann.lecun.com/exdb/mnist/'</span></span><br><span class="line">  file_names = [<span class="string">'train-images-idx3-ubyte.gz'</span>,</span><br><span class="line">                <span class="string">'train-labels-idx1-ubyte.gz'</span>,</span><br><span class="line">                <span class="string">'t10k-images-idx3-ubyte.gz'</span>,</span><br><span class="line">                <span class="string">'t10k-labels-idx1-ubyte.gz'</span>]</span><br><span class="line">  <span class="keyword">for</span> file_name <span class="keyword">in</span> file_names:</span><br><span class="line">    url = (url_base+file_name).format(**locals())</span><br><span class="line">    print(url)</span><br><span class="line">    out_path = os.path.join(data_dir,file_name)</span><br><span class="line">    cmd = [<span class="string">'curl'</span>, url, <span class="string">'-o'</span>, out_path]</span><br><span class="line">    print(<span class="string">'Downloading '</span>, file_name)</span><br><span class="line">    subprocess.call(cmd)</span><br><span class="line">    cmd = [<span class="string">'gzip'</span>, <span class="string">'-d'</span>, out_path]</span><br><span class="line">    print(<span class="string">'Decompressing '</span>, file_name)</span><br><span class="line">    subprocess.call(cmd)</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">prepare_data_dir</span><span class="params">(path = <span class="string">'./data'</span>)</span>:</span></span><br><span class="line">  <span class="keyword">if</span> <span class="keyword">not</span> os.path.exists(path):</span><br><span class="line">    os.mkdir(path)</span><br><span class="line"></span><br><span class="line"><span class="keyword">if</span> __name__ == <span class="string">'__main__'</span>:</span><br><span class="line">  args = parser.parse_args()</span><br><span class="line">  prepare_data_dir()</span><br><span class="line"></span><br><span class="line">  <span class="comment"># 如果datasets参数是 ['CelebA', 'celebA', 'celebA'] 其中之一</span></span><br><span class="line">  <span class="keyword">if</span> any(name <span class="keyword">in</span> args.datasets <span class="keyword">
187for</span> name <span class="keyword">in</span> [<span class="string">'CelebA'</span>, <span class="string">'celebA'</span>, <span class="string">'celebA'</span>]):</span><br><span class="line">    download_celeb_a(<span class="string">'./data'</span>)</span><br><span class="line">  <span class="keyword">if</span> <span class="string">'lsun'</span> <span class="keyword">in</span> args.datasets:</span><br><span class="line">    download_lsun(<span class="string">'./data'</span>)</span><br><span class="line">  <span class="keyword">if</span> <span class="string">'mnist'</span> <span class="keyword">in</span> args.datasets:</span><br><span class="line">    download_mnist(<span class="string">'./data'</span>)</span><br></pre></td></tr></table></figure>
188</li>
189<li><p>首先需要导入的包中,<strong>gzip</strong>和<strong>zipfile</strong>用于文件压缩和解压缩相关;<strong>argparse</strong>用于构建命令行参数;<strong>requests</strong>用于http请求下载网络文件资源;<strong>subprocess</strong>用于运行shell命令;<strong>tqdm</strong>用于进度条显示;<strong>six</strong>包用于python2和python3的兼容,比如 <span class="lang:default decode:true  crayon-inline ">from six.moves import urllib</span> 这句就是导入python2.x的urllib库。</p>
190</li>
191<li>上面的代码除了原作者加的注释之外,我也已经加了一部分注释,意思应该比较好理解了。主要做的事情,就是利用requests库从网络上将<strong>mnist</strong>,<strong>lsun</strong>以及<strong>celebA</strong>这三个数据集下载下来,保存在data目录下。注意<strong>mnist</strong>和<strong>celebA</strong>数据集下载下来之后还进行了解压缩。</li>
192<li>上面的三个数据集,<strong>mnist</strong>是著名的手写数字数据库,大家应该都已经很熟悉了;lsun是大型场景理解数据集(large-scale-scene-understanding);celebA是一个开源的人脸数据库。除了mnist之外,其余两个数据集体积都较大,celebA大概有20w+的图像,压缩文件体积为1.4G;而lsun有很多个场景不同的数据集,如果按照上面的脚本下载,下载的文件为bedroom数据集,压缩文件有46G之大,而且其实下载下来的文件解压后为mdb(Access数据库)格式,不是原始图片格式,不方便处理。所以我们实际会下载其他的数据集作为替代,比如这个<a href="http://lsun.cs.princeton.edu/challenge/2015/roomlayout/data/image.zip" target="_blank" rel="noopener">room layout estimation(2G)</a>数据。如果使用download.py脚本下载速度较慢的话,可以自行下载好数据集,然后放在data目录下即可。</li>
193</ul>
194<h4 id="2-2-main-py"><a href="#2-2-main-py" class="headerlink" title="2.2 main.py"></a>2.2 <strong>main.py</strong></h4><h4 id="nbsp-nbsp-nbsp-nbsp-main-py代码如下"><a href="#nbsp-nbsp-nbsp-nbsp-main-py代码如下" class="headerlink" title="&nbsp;&nbsp;&nbsp;&nbsp;main.py代码如下:"></a>&nbsp;&nbsp;&nbsp;&nbsp;main.py代码如下:</h4><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br><span class="line">44</span><br><span class="line">45</span><br><span class="line">46</span><br><span class="line">47</span><br><span class="line">48</span><br><span class="line">49</span><br><span class="line">50</span><br><span class="line">51</span><br><span class="line">52</span><br><span class="line">
19453</span><br><span class="line">54</span><br><span class="line">55</span><br><span class="line">56</span><br><span class="line">57</span><br><span class="line">58</span><br><span class="line">59</span><br><span class="line">60</span><br><span class="line">61</span><br><span class="line">62</span><br><span class="line">63</span><br><span class="line">64</span><br><span class="line">65</span><br><span class="line">66</span><br><span class="line">67</span><br><span class="line">68</span><br><span class="line">69</span><br><span class="line">70</span><br><span class="line">71</span><br><span class="line">72</span><br><span class="line">73</span><br><span class="line">74</span><br><span class="line">75</span><br><span class="line">76</span><br><span class="line">77</span><br><span class="line">78</span><br><span class="line">79</span><br><span class="line">80</span><br><span class="line">81</span><br><span class="line">82</span><br><span class="line">83</span><br><span class="line">84</span><br><span class="line">85</span><br><span class="line">86</span><br><span class="line">87</span><br><span class="line">88</span><br><span class="line">89</span><br><span class="line">90</span><br><span class="line">91</span><br><span class="line">92</span><br><span class="line">93</span><br><span class="line">94</span><br><span class="line">95</span><br><span class="line">96</span><br><span class="line">97</span><br><span class="line">98</span><br><span class="line">99</span><br><span class="line">100</span><br><span class="line">101</span><br><span class="line">102</span><br><span class="line">103</span><br><span class="line">104</span><br><span class="line">105</span><br><span class="line">106</span><br><span class="line">107</span><br><span class="line">108</span><br><span class="line">109</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">import</span> os</span><br><span class="line"><span class="keyword">import</span> scipy.misc</span><br><span class="line"><span class="keyword">import</span> numpy <span class="keyword">as</span> np</span><br><span class="line"></span><br><span class="line"><span class="keyword">from</span> model <span class="keyword">import</span> DCGAN</span><br><span class="line"><span class="keyword">from</span> utils <span class="keyword">import</span> pp, visualize, to_json, show_all_variables</span><br><span class="line"></span><br><span class="line"><span class="keyword">import</span> tensorflow <span class="keyword">as</span> tf</span><br><span class="line"></span><br><span class="line"><span class="comment"># tensorflow 定义命令行参数</span></span><br><span class="line">flags = tf.app.flags</span><br><span class="line"><span class="comment"># flag_name, default_value, docstring</span></span><br><span class="line">flags.DEFINE_integer(<span class="string">"epoch"</span>, <span class="number">25</span>, <span class="string">"Epoch to train [25]"</span>)</span><br><span class="line">flags.DEFINE_float(<span class="string">"learning_rate"</span>, <span class="number">0.0002</span>, <span class="string">"Learning rate of for adam [0.0002]"</span>)</span><br><span class="line">flags.DEFINE_float(<span class="string">"beta1"</span>, <span class="number">0.5</span>, <span class="string">"Momentum term of adam [0.5]"</span>)</span><br><span class="line">flags.DEFINE_float(<span class="string">"train_size"</span>, np.inf, <span class="string">"The size of train images [np.inf]"</span>)</span><br><span class="line">flags.DEFINE_integer(<span class="string">"batch_size"</span>, <span class="number">64</span>, <span class="string">"The size of batch images [64]"</span>)</span><br><span class="line">flags.DEFINE_integer(<span class="string">"input_height"</span>, <span class="number">108</span>, <span class="string">"The size of image to use (will be center cropped). [108]"</span>)</span><br><span class="line">flags.DEFINE_integer(<span class="string">"input_width"</span>, <span class="keyword">None</span>, <span class="string">"The size of image to use (will be center cropped). If None, same value as input_height [None]"</span>)</span><br><span class="line">flags.DEFINE_integer(<span class="string">"output_height"</span>, <span class="number">64</span>, <span class="string">"The size of the output images to produce [64]"</span>)</span><br><span class="line">flags.DEFINE_integer(<span class="string">"output_width"</span>, <span class="keyword">None</span>, <span class="string">"The size of the output images to produce. If None, same value as output_height [None]"</span>)</span><br><span class="line">flags.DEFINE_integer(<span class="string">"print_every"</span>,<span class="number">100</span>,<span class="string">"print train info every 100 iterations"</span>)</span><br><span class="line">flags.DEFINE_integer(<span class="string">"checkpoint_every"</span>,<span class="number">500</span>,<span class="string">"save checkpoint file every 500 iterations"</span>)</span><br><span class="line">flags.DEFINE_string(<span class="string">"dataset"</span>, <span class="string">
194"celebA"</span>, <span class="string">"The name of dataset [celebA, mnist, lsun]"</span>)</span><br><span class="line">flags.DEFINE_string(<span class="string">"input_fname_pattern"</span>, <span class="string">"*.jpg"</span>, <span class="string">"Glob pattern of filename of input images [*]"</span>)</span><br><span class="line">flags.DEFINE_string(<span class="string">"checkpoint_dir"</span>, <span class="string">"checkpoint"</span>, <span class="string">"Directory name to save the checkpoints [checkpoint]"</span>)</span><br><span class="line">flags.DEFINE_string(<span class="string">"data_dir"</span>, <span class="string">"./data"</span>, <span class="string">"Root directory of dataset [data]"</span>)</span><br><span class="line">flags.DEFINE_string(<span class="string">"sample_dir"</span>, <span class="string">"samples"</span>, <span class="string">"Directory name to save the image samples [samples]"</span>)</span><br><span class="line">flags.DEFINE_boolean(<span class="string">"train"</span>, <span class="keyword">False</span>, <span class="string">"True for training, False for testing [False]"</span>)</span><br><span class="line">flags.DEFINE_boolean(<span class="string">"crop"</span>, <span class="keyword">False</span>, <span class="string">"True for training, False for testing [False]"</span>)</span><br><span class="line">flags.DEFINE_boolean(<span class="string">"visualize"</span>, <span class="keyword">False</span>, <span class="string">"True for visualizing, False for nothing [False]"</span>)</span><br><span class="line">flags.DEFINE_integer(<span class="string">"generate_test_images"</span>, <span class="number">100</span>, <span class="string">"Number of images to generate during test. [100]"</span>)</span><br><span class="line">FLAGS = flags.FLAGS</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">main</span><span class="params">(_)</span>:</span></span><br><span class="line">  pp.pprint(flags.FLAGS.__flags)</span><br><span class="line"></span><br><span class="line">  <span class="comment"># 如果宽度没有指定,那么和高度一样</span></span><br><span class="line">  <span class="keyword">if</span> FLAGS.input_width <span class="keyword">is</span> <span class="keyword">None</span>:</span><br><span class="line">    FLAGS.input_width = FLAGS.input_height</span><br><span class="line">  <span class="keyword">if</span> FLAGS.output_width <span class="keyword">is</span> <span class="keyword">None</span>:</span><br><span class="line">    FLAGS.output_width = FLAGS.output_height</span><br><span class="line"></span><br><span class="line">  <span class="keyword">if</span> <span class="keyword">not</span> os.path.exists(FLAGS.checkpoint_dir):</span><br><span class="line">    os.makedirs(FLAGS.checkpoint_dir)</span><br><span class="line">  <span class="keyword">if</span> <span class="keyword">not</span> os.path.exists(FLAGS.sample_dir):</span><br><span class="line">    os.makedirs(FLAGS.sample_dir)</span><br><span class="line"></span><br><span class="line">  <span class="comment">#gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)</span></span><br><span class="line">  run_config = tf.ConfigProto()</span><br><span class="line">  run_config.gpu_options.allow_growth=<span class="keyword">True</span></span><br><span class="line"></span><br><span class="line">  <span class="keyword">with</span> tf.Session(config=run_config) <span class="keyword">as</span> sess:</span><br><span class="line">    <span class="keyword">if</span> FLAGS.dataset == <span class="string">'mnist'</span>:</span><br><span class="line">      dcgan = DCGAN(</span><br><span class="line">          sess,</span><br><span class="line">          input_width=FLAGS.input_width,</span><br><span class="line">          input_height=FLAGS.input_height,</span><br><span class="line">          output_width=FLAGS.output_width,</span><br><span class="line">          output_height=FLAGS.output_height,</span><br><span class="line">          batch_size=FLAGS.batch_size,</span><br><span class="line">          sample_num=FLAGS.batch_size,</span><br><span class="line">          y_dim=<span class="number">10</span>,</span><br><span class="line">          z_dim=FLAGS.generate_test_images,</span><br><span class="line">
194          dataset_name=FLAGS.dataset,</span><br><span class="line">          input_fname_pattern=FLAGS.input_fname_pattern,</span><br><span class="line">          crop=FLAGS.crop,</span><br><span class="line">          checkpoint_dir=FLAGS.checkpoint_dir,</span><br><span class="line">          sample_dir=FLAGS.sample_dir,</span><br><span class="line">          data_dir=FLAGS.data_dir)</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      dcgan = DCGAN(</span><br><span class="line">          sess,</span><br><span class="line">          input_width=FLAGS.input_width,</span><br><span class="line">          input_height=FLAGS.input_height,</span><br><span class="line">          output_width=FLAGS.output_width,</span><br><span class="line">          output_height=FLAGS.output_height,</span><br><span class="line">          batch_size=FLAGS.batch_size,</span><br><span class="line">          sample_num=FLAGS.batch_size,</span><br><span class="line">          z_dim=FLAGS.generate_test_images,</span><br><span class="line">          dataset_name=FLAGS.dataset,</span><br><span class="line">          input_fname_pattern=FLAGS.input_fname_pattern,</span><br><span class="line">          crop=FLAGS.crop,</span><br><span class="line">          checkpoint_dir=FLAGS.checkpoint_dir,</span><br><span class="line">          sample_dir=FLAGS.sample_dir,</span><br><span class="line">          data_dir=FLAGS.data_dir)</span><br><span class="line"></span><br><span class="line">    show_all_variables()</span><br><span class="line"></span><br><span class="line">    <span class="keyword">if</span> FLAGS.train:</span><br><span class="line">      dcgan.train(FLAGS)</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">        <span class="comment"># dcgan.load return:True,counter</span></span><br><span class="line">      <span class="keyword">if</span> <span class="keyword">not</span> dcgan.load(FLAGS.checkpoint_dir)[<span class="number">0</span>]: <span class="comment">#没有成功加载checkpoint file</span></span><br><span class="line">        <span class="keyword">raise</span> Exception(<span class="string">"[!] Train a model first, then run test mode"</span>)</span><br><span class="line"></span><br><span class="line"></span><br><span class="line">    <span class="comment"># to_json("./web/js/layers.js", [dcgan.h0_w, dcgan.h0_b, dcgan.g_bn0],</span></span><br><span class="line">    <span class="comment">#                 [dcgan.h1_w, dcgan.h1_b, dcgan.g_bn1],</span></span><br><span class="line">    <span class="comment">#                 [dcgan.h2_w, dcgan.h2_b, dcgan.g_bn2],</span></span><br><span class="line">    <span class="comment">#                 [dcgan.h3_w, dcgan.h3_b, dcgan.g_bn3],</span></span><br><span class="line">    <span class="comment">#                 [dcgan.h4_w, dcgan.h4_b, None])</span></span><br><span class="line"></span><br><span class="line">    <span class="comment"># Below is codes for visualization</span></span><br><span class="line">    OPTION = <span class="number">4</span></span><br><span class="line">    visualize(sess, dcgan, FLAGS, OPTION)</span><br><span class="line"></span><br><span class="line"><span class="keyword">if</span> __name__ == <span class="string">'__main__'</span>:</span><br><span class="line">  tf.app.run()</span><br></pre></td></tr></table></figure>
195<ul>
196<li>这里需要注意的是  <span class="lang:default decode:true    crayon-inline ">flags = tf.app.flags</span>  用于tensorflow构建命令行参数, <span class="lang:default decode:true    crayon-inline ">flags.DEFINE_xxx(param,default,description)</span> 用于定义命令行参数及其取值,第一个参数<strong>param</strong>是具体参数值,第二个参数<strong>default</strong>是参数默认取值,第三个参数<strong>description</strong>是参数描述字符串。</li>
197<li>在构建了sess之后,我们需要区分数据集是mnist还是其他数据集。因为mnist比较特殊,它有10个类别的数字图像,所以我们在构建<strong>DCGAN</strong>的时候需要额外多传递一个<strong>y_dim=10</strong>参数。 <span class="lang:python decode:true    crayon-inline ">show_all_variables</span> 函数用于显示model所有变量的具体信息。</li>
198<li>
198接下来如果是训练状态( <span class="lang:default decode:true    crayon-inline ">FLAGS.train == True</span> ),则进行模型训练( <span class="lang:default decode:true    crayon-inline ">dcgan.train(FLAGS)</span> ;否则进行测试,即加载之前训练时候保存的checkpoint文件,然后调用 <span class="lang:default decode:true    crayon-inline ">visualize</span> 函数进行test(该函数可以生成image或者gif,可视化展示训练的效果)。</li>
199<li><span class="lang:default decode:true    crayon-inline ">tf.app.run()</span> 是常用的tensorflow运行的起始命令。</li>
200</ul>
201<h4 id="2-3-model-py"><a href="#2-3-model-py" class="headerlink" title="2.3 model.py"></a>2.3 <strong>model.py</strong></h4><h4 id="nbsp-nbsp-nbsp-nbsp-model-py代码如下"><a href="#nbsp-nbsp-nbsp-nbsp-model-py代码如下" class="headerlink" title="&nbsp;&nbsp;&nbsp;&nbsp;model.py代码如下:"></a>&nbsp;&nbsp;&nbsp;&nbsp;model.py代码如下:</h4><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br><span class="line">44</span><br><span class="line">45</span><br><span class="line">46</span><br><span class="line">47</span><br><span class="line">48</span><br><span class="line">49</span><br><span class="line">50</span><br><span class="line">51</span><br><span class="line">52</span><br><span class="line">
20153</span><br><span class="line">54</span><br><span class="line">55</span><br><span class="line">56</span><br><span class="line">57</span><br><span class="line">58</span><br><span class="line">59</span><br><span class="line">60</span><br><span class="line">61</span><br><span class="line">62</span><br><span class="line">63</span><br><span class="line">64</span><br><span class="line">65</span><br><span class="line">66</span><br><span class="line">67</span><br><span class="line">68</span><br><span class="line">69</span><br><span class="line">70</span><br><span class="line">71</span><br><span class="line">72</span><br><span class="line">73</span><br><span class="line">74</span><br><span class="line">75</span><br><span class="line">76</span><br><span class="line">77</span><br><span class="line">78</span><br><span class="line">79</span><br><span class="line">80</span><br><span class="line">81</span><br><span class="line">82</span><br><span class="line">83</span><br><span class="line">84</span><br><span class="line">85</span><br><span class="line">86</span><br><span class="line">87</span><br><span class="line">88</span><br><span class="line">89</span><br><span class="line">90</span><br><span class="line">91</span><br><span class="line">92</span><br><span class="line">93</span><br><span class="line">94</span><br><span class="line">95</span><br><span class="line">96</span><br><span class="line">97</span><br><span class="line">98</span><br><span class="line">99</span><br><span class="line">100</span><br><span class="line">101</span><br><span class="line">102</span><br><span class="line">103</span><br><span class="line">104</span><br><span class="line">105</span><br><span class="line">106</span><br><span class="line">107</span><br><span class="line">108</span><br><span class="line">109</span><br><span class="line">110</span><br><span class="line">111</span><br><span class="line">112</span><br><span class="line">113</span><br><span class="line">114</span><br><span class="line">115</span><br><span class="line">116</span><br><span class="line">117</span><br><span class="line">118</span><br><span class="line">119</span><br><span class="line">120</span><br><span class="line">121</span><br><span class="line">122</span><br><span class="line">123</span><br><span class="line">124</span><br><span class="line">125</span><br><span class="line">126</span><br><span class="line">127</span><br><span class="line">128</span><br><span class="line">129</span><br><span class="line">130</span><br><span class="line">131</span><br><span class="line">132</span><br><span class="line">133</span><br><span class="line">134</span><br><span class="line">135</span><br><span class="line">136</span><br><span class="line">137</span><br><span class="line">138</span><br><span class="line">139</span><br><span class="line">140</span><br><span class="line">141</span><br><span class="line">142</span><br><span class="line">143</span><br><span class="line">144</span><br><span class="line">145</span><br><span class="line">146</span><br><span class="line">147</span><br><span class="line">148</span><br><span class="line">149</span><br><span class="line">150</span><br><span class="line">151</span><br><span class="line">152</span><br><span class="line">153</span><br><span class="line">154</span><br><span class="line">155</span><br><span class="line">156</span><br><span class="line">157</span><br><span class="line">158</span><br><span class="line">159</span><br><span class="line">160</span><br><span class="line">161</span><br><span class="line">162</span><br><span class="line">163</span><br><span class="line">164</span><br><span class="line">165</span><br><span class="line">166</span><br><span class="line">167</span><br><span class="line">168</span><br><span class="line">169</span><br><span class="line">170</span><br><span class="line">171</span><br><span class="line">172</span><br><span class="line">173</span><br><span class="line">174</span><br><span class="line">175</span><br><span class="line">176</span><br><span class="line">177</span><br><span class="line">178</span><br><span class="line">179</span><br><span class="line">180</span><br><span class="line">181</span><br><span class="line">182</span><br><span class="line">183</span><br><span class="line">184</span><br><span class="line">185</span><br><span class="line">186</span><br><span class="line">187</span><br><span class="line">188</span><br><span class="line">189</span><br><span class="line">190</span><br><span class="line">191</span><br><span class="line">192</span><br><span class="line">193</span><br><span class="line">
201194</span><br><span class="line">195</span><br><span class="line">196</span><br><span class="line">197</span><br><span class="line">198</span><br><span class="line">199</span><br><span class="line">200</span><br><span class="line">201</span><br><span class="line">202</span><br><span class="line">203</span><br><span class="line">204</span><br><span class="line">205</span><br><span class="line">206</span><br><span class="line">207</span><br><span class="line">208</span><br><span class="line">209</span><br><span class="line">210</span><br><span class="line">211</span><br><span class="line">212</span><br><span class="line">213</span><br><span class="line">214</span><br><span class="line">215</span><br><span class="line">216</span><br><span class="line">217</span><br><span class="line">218</span><br><span class="line">219</span><br><span class="line">220</span><br><span class="line">221</span><br><span class="line">222</span><br><span class="line">223</span><br><span class="line">224</span><br><span class="line">225</span><br><span class="line">226</span><br><span class="line">227</span><br><span class="line">228</span><br><span class="line">229</span><br><span class="line">230</span><br><span class="line">231</span><br><span class="line">232</span><br><span class="line">233</span><br><span class="line">234</span><br><span class="line">235</span><br><span class="line">236</span><br><span class="line">237</span><br><span class="line">238</span><br><span class="line">239</span><br><span class="line">240</span><br><span class="line">241</span><br><span class="line">242</span><br><span class="line">243</span><br><span class="line">244</span><br><span class="line">245</span><br><span class="line">246</span><br><span class="line">247</span><br><span class="line">248</span><br><span class="line">249</span><br><span class="line">250</span><br><span class="line">251</span><br><span class="line">252</span><br><span class="line">253</span><br><span class="line">254</span><br><span class="line">255</span><br><span class="line">256</span><br><span class="line">257</span><br><span class="line">258</span><br><span class="line">259</span><br><span class="line">260</span><br><span class="line">261</span><br><span class="line">262</span><br><span class="line">263</span><br><span class="line">264</span><br><span class="line">265</span><br><span class="line">266</span><br><span class="line">267</span><br><span class="line">268</span><br><span class="line">269</span><br><span class="line">270</span><br><span class="line">271</span><br><span class="line">272</span><br><span class="line">273</span><br><span class="line">274</span><br><span class="line">275</span><br><span class="line">276</span><br><span class="line">277</span><br><span class="line">278</span><br><span class="line">279</span><br><span class="line">280</span><br><span class="line">281</span><br><span class="line">282</span><br><span class="line">283</span><br><span class="line">284</span><br><span class="line">285</span><br><span class="line">286</span><br><span class="line">287</span><br><span class="line">288</span><br><span class="line">289</span><br><span class="line">290</span><br><span class="line">291</span><br><span class="line">292</span><br><span class="line">293</span><br><span class="line">294</span><br><span class="line">295</span><br><span class="line">296</span><br><span class="line">297</span><br><span class="line">298</span><br><span class="line">299</span><br><span class="line">300</span><br><span class="line">301</span><br><span class="line">302</span><br><span class="line">303</span><br><span class="line">304</span><br><span class="line">305</span><br><span class="line">306</span><br><span class="line">307</span><br><span class="line">308</span><br><span class="line">309</span><br><span class="line">310</span><br><span class="line">311</span><br><span class="line">312</span><br><span class="line">313</span><br><span class="line">314</span><br><span class="line">315</span><br><span class="line">316</span><br><span class="line">317</span><br><span class="line">318</span><br><span class="line">319</span><br><span class="line">320</span><br><span class="line">321</span><br><span class="line">322</span><br><span class="line">323</span><br><span class="line">324</span><br><span class="line">325</span><br><span class="line">326</span><br><span class="line">327</span><br><span class="line">
201328</span><br><span class="line">329</span><br><span class="line">330</span><br><span class="line">331</span><br><span class="line">332</span><br><span class="line">333</span><br><span class="line">334</span><br><span class="line">335</span><br><span class="line">336</span><br><span class="line">337</span><br><span class="line">338</span><br><span class="line">339</span><br><span class="line">340</span><br><span class="line">341</span><br><span class="line">342</span><br><span class="line">343</span><br><span class="line">344</span><br><span class="line">345</span><br><span class="line">346</span><br><span class="line">347</span><br><span class="line">348</span><br><span class="line">349</span><br><span class="line">350</span><br><span class="line">351</span><br><span class="line">352</span><br><span class="line">353</span><br><span class="line">354</span><br><span class="line">355</span><br><span class="line">356</span><br><span class="line">357</span><br><span class="line">358</span><br><span class="line">359</span><br><span class="line">360</span><br><span class="line">361</span><br><span class="line">362</span><br><span class="line">363</span><br><span class="line">364</span><br><span class="line">365</span><br><span class="line">366</span><br><span class="line">367</span><br><span class="line">368</span><br><span class="line">369</span><br><span class="line">370</span><br><span class="line">371</span><br><span class="line">372</span><br><span class="line">373</span><br><span class="line">374</span><br><span class="line">375</span><br><span class="line">376</span><br><span class="line">377</span><br><span class="line">378</span><br><span class="line">379</span><br><span class="line">380</span><br><span class="line">381</span><br><span class="line">382</span><br><span class="line">383</span><br><span class="line">384</span><br><span class="line">385</span><br><span class="line">386</span><br><span class="line">387</span><br><span class="line">388</span><br><span class="line">389</span><br><span class="line">390</span><br><span class="line">391</span><br><span class="line">392</span><br><span class="line">393</span><br><span class="line">394</span><br><span class="line">395</span><br><span class="line">396</span><br><span class="line">397</span><br><span class="line">398</span><br><span class="line">399</span><br><span class="line">400</span><br><span class="line">401</span><br><span class="line">402</span><br><span class="line">403</span><br><span class="line">404</span><br><span class="line">405</span><br><span class="line">406</span><br><span class="line">407</span><br><span class="line">408</span><br><span class="line">409</span><br><span class="line">410</span><br><span class="line">411</span><br><span class="line">412</span><br><span class="line">413</span><br><span class="line">414</span><br><span class="line">415</span><br><span class="line">416</span><br><span class="line">417</span><br><span class="line">418</span><br><span class="line">419</span><br><span class="line">420</span><br><span class="line">421</span><br><span class="line">422</span><br><span class="line">423</span><br><span class="line">
201424</span><br><span class="line">425</span><br><span class="line">426</span><br><span class="line">427</span><br><span class="line">428</span><br><span class="line">429</span><br><span class="line">430</span><br><span class="line">431</span><br><span class="line">432</span><br><span class="line">433</span><br><span class="line">434</span><br><span class="line">435</span><br><span class="line">436</span><br><span class="line">437</span><br><span class="line">438</span><br><span class="line">439</span><br><span class="line">440</span><br><span class="line">441</span><br><span class="line">442</span><br><span class="line">443</span><br><span class="line">444</span><br><span class="line">445</span><br><span class="line">446</span><br><span class="line">447</span><br><span class="line">448</span><br><span class="line">449</span><br><span class="line">450</span><br><span class="line">451</span><br><span class="line">452</span><br><span class="line">453</span><br><span class="line">454</span><br><span class="line">455</span><br><span class="line">456</span><br><span class="line">457</span><br><span class="line">458</span><br><span class="line">459</span><br><span class="line">460</span><br><span class="line">461</span><br><span class="line">462</span><br><span class="line">463</span><br><span class="line">464</span><br><span class="line">465</span><br><span class="line">466</span><br><span class="line">467</span><br><span class="line">468</span><br><span class="line">469</span><br><span class="line">470</span><br><span class="line">471</span><br><span class="line">472</span><br><span class="line">473</span><br><span class="line">474</span><br><span class="line">475</span><br><span class="line">476</span><br><span class="line">477</span><br><span class="line">478</span><br><span class="line">479</span><br><span class="line">480</span><br><span class="line">481</span><br><span class="line">482</span><br><span class="line">483</span><br><span class="line">484</span><br><span class="line">485</span><br><span class="line">486</span><br><span class="line">487</span><br><span class="line">488</span><br><span class="line">489</span><br><span class="line">490</span><br><span class="line">491</span><br><span class="line">492</span><br><span class="line">493</span><br><span class="line">494</span><br><span class="line">495</span><br><span class="line">496</span><br><span class="line">497</span><br><span class="line">498</span><br><span class="line">499</span><br><span class="line">500</span><br><span class="line">501</span><br><span class="line">502</span><br><span class="line">503</span><br><span class="line">504</span><br><span class="line">505</span><br><span class="line">506</span><br><span class="line">507</span><br><span class="line">508</span><br><span class="line">509</span><br><span class="line">510</span><br><span class="line">511</span><br><span class="line">512</span><br><span class="line">513</span><br><span class="line">514</span><br><span class="line">515</span><br><span class="line">516</span><br><span class="line">517</span><br><span class="line">518</span><br><span class="line">519</span><br><span class="line">520</span><br><span class="line">521</span><br><span class="line">522</span><br><span class="line">523</span><br><span class="line">524</span><br><span class="line">525</span><br><span class="line">526</span><br><span class="line">527</span><br><span class="line">528</span><br><span class="line">529</span><br><span class="line">530</span><br><span class="line">531</span><br><span class="line">532</span><br><span class="line">533</span><br><span class="line">534</span><br><span class="line">535</span><br><span class="line">536</span><br><span class="line">537</span><br><span class="line">538</span><br><span class="line">539</span><br><span class="line">540</span><br><span class="line">541</span><br><span class="line">542</span><br><span class="line">543</span><br><span class="line">544</span><br><span class="line">545</span><br><span class="line">546</span><br><span class="line">547</span><br><span class="line">548</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">from</span> __future__ <span class="keyword">
201import</span> division</span><br><span class="line"><span class="keyword">import</span> os</span><br><span class="line"><span class="keyword">import</span> time</span><br><span class="line"><span class="keyword">import</span> math</span><br><span class="line"><span class="keyword">from</span> glob <span class="keyword">import</span> glob</span><br><span class="line"><span class="keyword">import</span> tensorflow <span class="keyword">as</span> tf</span><br><span class="line"><span class="keyword">import</span> numpy <span class="keyword">as</span> np</span><br><span class="line"><span class="keyword">from</span> six.moves <span class="keyword">import</span> xrange</span><br><span class="line"></span><br><span class="line"><span class="keyword">from</span> ops <span class="keyword">import</span> *</span><br><span class="line"><span class="keyword">from</span> utils <span class="keyword">import</span> *</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">conv_out_size_same</span><span class="params">(size, stride)</span>:</span></span><br><span class="line">  <span class="keyword">return</span> int(math.ceil(float(size) / float(stride)))</span><br><span class="line"></span><br><span class="line"><span class="class"><span class="keyword">class</span> <span class="title">DCGAN</span><span class="params">(object)</span>:</span></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">__init__</span><span class="params">(self, sess, input_height=<span class="number">108</span>, input_width=<span class="number">108</span>, crop=True,</span></span></span><br><span class="line"><span class="function"><span class="params">         batch_size=<span class="number">64</span>, sample_num = <span class="number">64</span>, output_height=<span class="number">64</span>, output_width=<span class="number">64</span>,</span></span></span><br><span class="line"><span class="function"><span class="params">         y_dim=None, z_dim=<span class="number">100</span>, gf_dim=<span class="number">64</span>, df_dim=<span class="number">64</span>,</span></span></span><br><span class="line"><span class="function"><span class="params">         gfc_dim=<span class="number">1024</span>, dfc_dim=<span class="number">1024</span>, c_dim=<span class="number">3</span>, dataset_name=<span class="string">'default'</span>,</span></span></span><br><span class="line"><span class="function"><span class="params">         input_fname_pattern=<span class="string">'*.jpg'</span>, checkpoint_dir=None, sample_dir=None, data_dir=<span class="string">'data'</span>)</span>:</span></span><br><span class="line">    <span class="string">"""</span></span><br><span class="line"><span class="string"></span></span><br><span class="line"><span class="string">    Args:</span></span><br><span class="line"><span class="string">      sess: TensorFlow session</span></span><br><span class="line"><span class="string">      batch_size: The size of batch. Should be specified before training.</span></span><br><span class="line"><span class="string">      y_dim: (optional) Dimension of dim for y. [None]</span></span><br><span class="line"><span class="string">      z_dim: (optional) Dimension of dim for Z. [100]</span></span><br><span class="line"><span class="string">      # 生成器第一个卷积层 filters size</span></span><br><span class="line"><span class="string">      gf_dim: (optional) Dimension of gen filters in first conv layer. [64]</span></span><br><span class="line"><span class="string">      # 鉴别器第一个卷积层filters size</span></span><br><span class="line"><span class="string">      df_dim: (optional) Dimension of discrim filters in first conv layer. [64]</span></span><br><span class="line"><span class="string">      # 生成器全连接层units size</span></span><br><span class="line"><span class="string">      gfc_dim: (optional) Dimension of gen units for for fully connected layer. [1024]</span></span><br><span class="line"><span class="string">      # 鉴别器全连接层units size</span></span><br><span class="line"><span class="string">      dfc_dim: (optional) Dimension of discrim units for fully connected layer. [1024]</span></span><br><span class="line"><span class="string">      # image channel</span></span><br><span class="line"><span class="string">      c_dim: (optional) Dimension of image color. For grayscale input, set to 1. [3]</span></span><br><span class="line"><span class="string">    """</span></span><br><span class="line">    self.sess = sess</span><br><span class="line">    self.crop = crop</span><br><span class="line"></span><br><span class="line">    self.batch_size = batch_size</span><br><span class="line">    self.sample_num = sample_num</span><br><span class="line"></span><br><span class="line">    self.input_height = input_height</span><br><span class="line">    self.input_width = input_width</span><br><span class="line">    self.output_height = output_height</span><br><span class="line">    self.output_width = output_width</span><br><span class="line"></span><br><span class="line">    self.y_dim = y_dim</span><br><span class="line">    self.z_dim = z_dim</span><br><span class="line"></span><br><span class="line">    self.gf_dim = gf_dim</span><br><span class="line">    self.df_dim = df_dim</span><br><span class="line"></span><br><span class="line">    self.gfc_dim = gfc_dim</span><br><span class="line">    self.dfc_dim = dfc_dim</span><br><span class="line"></span><br><span class="line">    <span class="comment"># batch normalization : deals with poor initialization helps gradient flow</span></span><br><span class="line">    self.d_bn1 = batch_norm(name=<span class="string">'d_bn1'</span>)</span><br><span class="line">    self.d_bn2 = batch_norm(name=<span class="string">'d_bn2'</span>)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">if</span> <span class="keyword">not</span> self.y_dim:</span><br><span class="line">      self.d_bn3 = batch_norm(name=<span class="string">'d_bn3'</span>)</span><br><span class="line"></span><br><span class="line">    self.g_bn0 = batch_norm(name=<span class="string">'g_bn0'</span>)</span><br><span class="line">    self.g_bn1 = batch_norm(name=<span class="string">'g_bn1'</span>)</span><br><span class="line">    self.g_bn2 = batch_norm(name=<span class="string">'g_bn2'</span>)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">if</span> <span class="keyword">not</span> self.y_dim:</span><br><span class="line">      self.g_bn3 = batch_norm(name=<span class="string">'g_bn3'</span>)</span><br><span class="line"></span><br><span class="line">    self.dataset_name = dataset_name</span><br><span class="line">    self.input_fname_pattern = input_fname_pattern</span><br><span class="line">    self.checkpoint_dir = checkpoint_dir</span><br><span class="line">    self.data_dir = data_dir</span><br><span class="line"></span><br><span class="line">    <span class="keyword">if</span> self.dataset_name == <span class="string">
201'mnist'</span>:</span><br><span class="line">      self.data_X, self.data_y = self.load_mnist()</span><br><span class="line">      self.c_dim = self.data_X[<span class="number">0</span>].shape[<span class="number">-1</span>]</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      <span class="comment"># dir *.jpg</span></span><br><span class="line">      self.data = glob(os.path.join(self.data_dir, self.dataset_name, self.input_fname_pattern))</span><br><span class="line">      imreadImg = imread(self.data[<span class="number">0</span>])</span><br><span class="line">      <span class="keyword">if</span> len(imreadImg.shape) &amp;gt;= <span class="number">3</span>: <span class="comment">#check if image is a non-grayscale image by checking channel number</span></span><br><span class="line">        self.c_dim = imread(self.data[<span class="number">0</span>]).shape[<span class="number">-1</span>] <span class="comment"># color image,get image channel</span></span><br><span class="line">      <span class="keyword">else</span>:</span><br><span class="line">        self.c_dim = <span class="number">1</span></span><br><span class="line"></span><br><span class="line">    self.grayscale = (self.c_dim == <span class="number">1</span>) <span class="comment"># 是否是灰度图像</span></span><br><span class="line"></span><br><span class="line">    self.build_model()</span><br><span class="line"></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">build_model</span><span class="params">(self)</span>:</span></span><br><span class="line">    <span class="keyword">if</span> self.y_dim:</span><br><span class="line">      self.y = tf.placeholder(tf.float32, [self.batch_size, self.y_dim], name=<span class="string">'y'</span>)</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      self.y = <span class="keyword">None</span></span><br><span class="line"></span><br><span class="line">    <span class="keyword">if</span> self.crop:</span><br><span class="line">      image_dims = [self.output_height, self.output_width, self.c_dim]</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      image_dims = [self.input_height, self.input_width, self.c_dim]</span><br><span class="line"></span><br><span class="line">    <span class="comment"># self.inputs shape:(batch_size,height,width,channel)</span></span><br><span class="line">    self.inputs = tf.placeholder(</span><br><span class="line">      tf.float32, [self.batch_size] + image_dims, name=<span class="string">'real_images'</span>)</span><br><span class="line"></span><br><span class="line">    inputs = self.inputs</span><br><span class="line"></span><br><span class="line">    self.z = tf.placeholder(</span><br><span class="line">      tf.float32, [<span class="keyword">None</span>, self.z_dim], name=<span class="string">'z'</span>)</span><br><span class="line">    <span class="comment"># 直方图可视化</span></span><br><span class="line">    self.z_sum = histogram_summary(<span class="string">"z"</span>, self.z)</span><br><span class="line"></span><br><span class="line">    self.G                  = self.generator(self.z, self.y)</span><br><span class="line">    self.D, self.D_logits   = self.discriminator(inputs, self.y, reuse=<span class="keyword">False</span>)</span><br><span class="line">    self.sampler            = self.sampler(self.z, self.y)</span><br><span class="line">    self.D_, self.D_logits_ = self.discriminator(self.G, self.y, reuse=<span class="keyword">True</span>)</span><br><span class="line">    </span><br><span class="line">    self.d_sum = histogram_summary(<span class="string">"d"</span>, self.D)</span><br><span class="line">    self.d__sum = histogram_summary(<span class="string">"d_"</span>, self.D_)</span><br><span class="line">    self.G_sum = image_summary(<span class="string">"G"</span>, self.G)</span><br><span class="line"></span><br><span class="line">    <span class="function"><span class="keyword">def</span> <span class="title">sigmoid_cross_entropy_with_logits</span><span class="params">(x, y)</span>:</span></span><br><span class="line">
201      <span class="keyword">try</span>:</span><br><span class="line">        <span class="keyword">return</span> tf.nn.sigmoid_cross_entropy_with_logits(logits=x, labels=y)</span><br><span class="line">      <span class="keyword">except</span>:</span><br><span class="line">        <span class="keyword">return</span> tf.nn.sigmoid_cross_entropy_with_logits(logits=x, targets=y)</span><br><span class="line"></span><br><span class="line">    self.d_loss_real = tf.reduce_mean(</span><br><span class="line">      sigmoid_cross_entropy_with_logits(self.D_logits, tf.ones_like(self.D)))</span><br><span class="line">    self.d_loss_fake = tf.reduce_mean(</span><br><span class="line">      sigmoid_cross_entropy_with_logits(self.D_logits_, tf.zeros_like(self.D_)))</span><br><span class="line">    self.g_loss = tf.reduce_mean(</span><br><span class="line">      sigmoid_cross_entropy_with_logits(self.D_logits_, tf.ones_like(self.D_)))</span><br><span class="line"></span><br><span class="line">    <span class="comment"># scalar_summary:Outputs a `Summary` protocol buffer containing a single scalar value</span></span><br><span class="line">    <span class="comment"># 返回一个scalar</span></span><br><span class="line">    self.d_loss_real_sum = scalar_summary(<span class="string">"d_loss_real"</span>, self.d_loss_real)</span><br><span class="line">    self.d_loss_fake_sum = scalar_summary(<span class="string">"d_loss_fake"</span>, self.d_loss_fake)</span><br><span class="line">                          </span><br><span class="line">    self.d_loss = self.d_loss_real + self.d_loss_fake</span><br><span class="line"></span><br><span class="line">    self.g_loss_sum = scalar_summary(<span class="string">"g_loss"</span>, self.g_loss)</span><br><span class="line">    self.d_loss_sum = scalar_summary(<span class="string">"d_loss"</span>, self.d_loss)</span><br><span class="line"></span><br><span class="line">    t_vars = tf.trainable_variables()</span><br><span class="line"></span><br><span class="line">    self.d_vars = [var <span class="keyword">for</span> var <span class="keyword">in</span> t_vars <span class="keyword">if</span> <span class="string">'d_'</span> <span class="keyword">in</span> var.name] <span class="comment"># 鉴别器相关变量</span></span><br><span class="line">    self.g_vars = [var <span class="keyword">for</span> var <span class="keyword">in</span> t_vars <span class="keyword">if</span> <span class="string">'g_'</span> <span class="keyword">in</span> var.name] <span class="comment"># 生成器相关变量</span></span><br><span class="line"></span><br><span class="line">    self.saver = tf.train.Saver()</span><br><span class="line"></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">train</span><span class="params">(self, config)</span>:</span></span><br><span class="line">    d_optim = tf.train.AdamOptimizer(config.learning_rate, beta1=config.beta1) \</span><br><span class="line">              .minimize(self.d_loss, var_list=self.d_vars)</span><br><span class="line">    g_optim = tf.train.AdamOptimizer(config.learning_rate, beta1=config.beta1) \</span><br><span class="line">              .minimize(self.g_loss, var_list=self.g_vars)</span><br><span class="line">    <span class="keyword">try</span>:</span><br><span class="line">      tf.global_variables_initializer().run()</span><br><span class="line">    <span class="keyword">except</span>:</span><br><span class="line">      tf.initialize_all_variables().run()</span><br><span class="line"></span><br><span class="line">    self.g_sum = merge_summary([self.z_sum, self.d__sum,</span><br><span class="line">      self.G_sum, self.d_loss_fake_sum, self.g_loss_sum])</span><br><span class="line">    self.d_sum = merge_summary(</span><br><span class="line">        [self.z_sum, self.d_sum, self.d_loss_real_sum, self.d_loss_sum])</span><br><span class="line">    self.writer = SummaryWriter(<span class="string">"./logs"</span>, self.sess.graph)</span><br><span class="line"></span><br><span class="line">    sample_z = np.random.uniform(<span class="number">-1</span>, <span class="number">1</span>, size=(self.sample_num , self.z_dim))</span><br><span class="line">    </span><br><span class="line">    <span class="keyword">if</span> config.dataset == <span class="string">
201'mnist'</span>:</span><br><span class="line">      sample_inputs = self.data_X[<span class="number">0</span>:self.sample_num]</span><br><span class="line">      sample_labels = self.data_y[<span class="number">0</span>:self.sample_num]</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      <span class="comment"># self.data is like:["0.jpg","1.jpg",...]</span></span><br><span class="line">      sample_files = self.data[<span class="number">0</span>:self.sample_num]</span><br><span class="line">      sample = [</span><br><span class="line">          <span class="comment"># get_image返回的是取值为(-1,1)的,shape为(resize_height,resize_width)的</span></span><br><span class="line">          <span class="comment"># ndarray</span></span><br><span class="line">          get_image(sample_file,</span><br><span class="line">                    input_height=self.input_height,</span><br><span class="line">                    input_width=self.input_width,</span><br><span class="line">                    resize_height=self.output_height,</span><br><span class="line">                    resize_width=self.output_width,</span><br><span class="line">                    crop=self.crop,</span><br><span class="line">                    grayscale=self.grayscale) <span class="keyword">for</span> sample_file <span class="keyword">in</span> sample_files]</span><br><span class="line">      <span class="keyword">if</span> (self.grayscale):</span><br><span class="line">        <span class="comment"># 灰度图像的channel为1</span></span><br><span class="line">        sample_inputs = np.array(sample).astype(np.float32)[:, :, :, <span class="keyword">None</span>]</span><br><span class="line">      <span class="keyword">else</span>:</span><br><span class="line">        <span class="comment"># color image</span></span><br><span class="line">        sample_inputs = np.array(sample).astype(np.float32)</span><br><span class="line">  </span><br><span class="line">    counter = <span class="number">1</span></span><br><span class="line">    start_time = time.time()</span><br><span class="line">    could_load, checkpoint_counter = self.load(self.checkpoint_dir)</span><br><span class="line">    <span class="keyword">if</span> could_load:</span><br><span class="line">      counter = checkpoint_counter</span><br><span class="line">      print(<span class="string">" [*] Load SUCCESS"</span>)</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      print(<span class="string">" [!] Load failed..."</span>)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">for</span> epoch <span class="keyword">in</span> xrange(config.epoch):</span><br><span class="line">      <span class="keyword">if</span> config.dataset == <span class="string">'mnist'</span>:</span><br><span class="line">        batch_idxs = min(len(self.data_X), config.train_size) // config.batch_size</span><br><span class="line">      <span class="keyword">else</span>:</span><br><span class="line">        <span class="comment"># self.data is like:["0.jpg","1.jpg",...]</span></span><br><span class="line">        self.data = glob(os.path.join(</span><br><span class="line">          config.data_dir, config.dataset, self.input_fname_pattern))</span><br><span class="line">        batch_idxs = min(len(self.data), config.train_size) // config.batch_size</span><br><span class="line"></span><br><span class="line">      <span class="keyword">for</span> idx <span class="keyword">in</span> xrange(<span class="number">0</span>, batch_idxs):</span><br><span class="line">        <span class="keyword">if</span> config.dataset == <span class="string">'mnist'</span>:</span><br><span class="line">          batch_images = self.data_X[idx*config.batch_size:(idx+<span class="number">1</span>)*config.batch_size]</span><br><span class="line">          batch_labels = self.data_y[idx*config.batch_size:(idx+<span class="number">1</span>)*config.batch_size]</span><br><span class="line">        <span class="keyword">else</span>:</span><br><span class="line">          batch_files = self.data[idx*config.batch_size:(idx+<span class="number">1</span>)*config.batch_size]</span><br><span class="line">          batch = [</span><br><span class="line">              get_image(batch_file,</span><br><span class="line">                        input_height=self.input_height,</span><br><span class="line">                        input_width=self.input_width,</span><br><span class="line">                        resize_height=self.output_height,</span><br><span class="line">                        resize_width=self.output_width,</span><br><span class="line">                        crop=self.crop,</span><br><span class="line">                        grayscale=self.grayscale) <span class="keyword">
201for</span> batch_file <span class="keyword">in</span> batch_files]</span><br><span class="line">          <span class="keyword">if</span> self.grayscale:</span><br><span class="line">            <span class="comment"># add a channel for grayscale</span></span><br><span class="line">            <span class="comment"># batch_images shape:(batch,height,width,channel)</span></span><br><span class="line">            batch_images = np.array(batch).astype(np.float32)[:, :, :, <span class="keyword">None</span>]</span><br><span class="line">          <span class="keyword">else</span>:</span><br><span class="line">            batch_images = np.array(batch).astype(np.float32)</span><br><span class="line">        <span class="comment"># add noise</span></span><br><span class="line">        batch_z = np.random.uniform(<span class="number">-1</span>, <span class="number">1</span>, [config.batch_size, self.z_dim]) \</span><br><span class="line">              .astype(np.float32)</span><br><span class="line"></span><br><span class="line">        <span class="keyword">if</span> config.dataset == <span class="string">'mnist'</span>:</span><br><span class="line">          <span class="comment"># Update D network</span></span><br><span class="line">          _, summary_str = self.sess.run([d_optim, self.d_sum],</span><br><span class="line">            feed_dict=&#123; </span><br><span class="line">              self.inputs: batch_images,</span><br><span class="line">              self.z: batch_z,</span><br><span class="line">              self.y:batch_labels,</span><br><span class="line">            &#125;)</span><br><span class="line">          <span class="comment"># 用于可视化</span></span><br><span class="line">          self.writer.add_summary(summary_str, counter)</span><br><span class="line"></span><br><span class="line">          <span class="comment"># Update G network</span></span><br><span class="line">          _, summary_str = self.sess.run([g_optim, self.g_sum],</span><br><span class="line">            feed_dict=&#123;</span><br><span class="line">              self.z: batch_z, </span><br><span class="line">              self.y:batch_labels,</span><br><span class="line">            &#125;)</span><br><span class="line">          self.writer.add_summary(summary_str, counter)</span><br><span class="line"></span><br><span class="line">          <span class="comment"># Run g_optim twice to make sure that d_loss does not go to zero (different from paper)</span></span><br><span class="line">          _, summary_str = self.sess.run([g_optim, self.g_sum],</span><br><span class="line">            feed_dict=&#123; self.z: batch_z, self.y:batch_labels &#125;)</span><br><span class="line">          self.writer.add_summary(summary_str, counter)</span><br><span class="line">          </span><br><span class="line">          errD_fake = self.d_loss_fake.eval(&#123;</span><br><span class="line">              self.z: batch_z, </span><br><span class="line">              self.y:batch_labels</span><br><span class="line">          &#125;)</span><br><span class="line">          errD_real = self.d_loss_real.eval(&#123;</span><br><span class="line">              self.inputs: batch_images,</span><br><span class="line">              self.y:batch_labels</span><br><span class="line">          &#125;)</span><br><span class="line">          errG = self.g_loss.eval(&#123;</span><br><span class="line">              self.z: batch_z,</span><br><span class="line">              self.y: batch_labels</span><br><span class="line">          &#125;)</span><br><span class="line">        <span class="keyword">else</span>:</span><br><span class="line">          <span class="comment"># Update D network</span></span><br><span class="line">          _, summary_str = self.sess.run([d_optim, self.d_sum],</span><br><span class="line">            feed_dict=&#123; self.inputs: batch_images, self.z: batch_z &#125;)</span><br><span class="line">          self.writer.add_summary(summary_str, counter)</span><br><span class="line"></span><br><span class="line">          <span class="comment"># Update G network</span></span><br><span class="line">          _, summary_str = self.sess.run([g_optim, self.g_sum],</span><br><span class="line">            feed_dict=&#123; self.z: batch_z &#125;)</span><br><span class="line">          self.writer.add_summary(summary_str, counter)</span><br><span class="line"></span><br><span class="line">
201          <span class="comment"># Run g_optim twice to make sure that d_loss does not go to zero (different from paper)</span></span><br><span class="line">          _, summary_str = self.sess.run([g_optim, self.g_sum],</span><br><span class="line">            feed_dict=&#123; self.z: batch_z &#125;)</span><br><span class="line">          self.writer.add_summary(summary_str, counter)</span><br><span class="line">          </span><br><span class="line">          errD_fake = self.d_loss_fake.eval(&#123; self.z: batch_z &#125;)</span><br><span class="line">          errD_real = self.d_loss_real.eval(&#123; self.inputs: batch_images &#125;)</span><br><span class="line">          errG = self.g_loss.eval(&#123;self.z: batch_z&#125;)</span><br><span class="line"></span><br><span class="line">        counter += <span class="number">1</span></span><br><span class="line">        print(<span class="string">"Epoch: [%2d/%2d] [%4d/%4d] time: %4.4f, d_loss: %.8f, g_loss: %.8f"</span> \</span><br><span class="line">          % (epoch, config.epoch, idx, batch_idxs,</span><br><span class="line">            time.time() - start_time, errD_fake+errD_real, errG))</span><br><span class="line">        <span class="comment"># np.mod:Return element-wise remainder of division.</span></span><br><span class="line">        <span class="comment"># 每100次生成一次samples</span></span><br><span class="line">        <span class="keyword">if</span> np.mod(counter, config.print_every) == <span class="number">1</span>:</span><br><span class="line">          <span class="keyword">if</span> config.dataset == <span class="string">'mnist'</span>:</span><br><span class="line">            samples, d_loss, g_loss = self.sess.run(</span><br><span class="line">              [self.sampler, self.d_loss, self.g_loss],</span><br><span class="line">              feed_dict=&#123;</span><br><span class="line">                  self.z: sample_z,</span><br><span class="line">                  self.inputs: sample_inputs,</span><br><span class="line">                  self.y:sample_labels,</span><br><span class="line">              &#125;</span><br><span class="line">            )</span><br><span class="line">            <span class="comment"># 保存生成的样本</span></span><br><span class="line">            save_images(samples, image_manifold_size(samples.shape[<span class="number">0</span>]),</span><br><span class="line">                  <span class="string">'./&#123;&#125;/train_&#123;:02d&#125;_&#123;:04d&#125;.png'</span>.format(config.sample_dir, epoch, idx))</span><br><span class="line">            print(<span class="string">"[Sample] d_loss: %.8f, g_loss: %.8f"</span> % (d_loss, g_loss)) </span><br><span class="line">          <span class="keyword">else</span>:</span><br><span class="line">            <span class="keyword">try</span>:</span><br><span class="line">              samples, d_loss, g_loss = self.sess.run(</span><br><span class="line">                [self.sampler, self.d_loss, self.g_loss],</span><br><span class="line">                feed_dict=&#123;</span><br><span class="line">                    self.z: sample_z,</span><br><span class="line">                    self.inputs: sample_inputs,</span><br><span class="line">                &#125;,</span><br><span class="line">              )</span><br><span class="line">              save_images(samples, image_manifold_size(samples.shape[<span class="number">0</span>]),</span><br><span class="line">                    <span class="string">'./&#123;&#125;/train_&#123;:02d&#125;_&#123;:04d&#125;.png'</span>.format(config.sample_dir, epoch, idx))</span><br><span class="line">              print(<span class="string">"[Sample] d_loss: %.8f, g_loss: %.8f"</span> % (d_loss, g_loss)) </span><br><span class="line">            <span class="keyword">except</span>:</span><br><span class="line">              print(<span class="string">"one pic error!..."</span>)</span><br><span class="line">        <span class="comment"># 每500次保存一下checkpoint</span></span><br><span class="line">        <span class="keyword">if</span>
201 np.mod(counter, config.checkpoint_every) == <span class="number">2</span>: <span class="comment"># save checkpoint file</span></span><br><span class="line">          self.save(config.checkpoint_dir, counter)</span><br><span class="line"></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">discriminator</span><span class="params">(self, image, y=None, reuse=False)</span>:</span></span><br><span class="line">    <span class="keyword">with</span> tf.variable_scope(<span class="string">"discriminator"</span>) <span class="keyword">as</span> scope:</span><br><span class="line">      <span class="keyword">if</span> reuse:</span><br><span class="line">        scope.reuse_variables()</span><br><span class="line"></span><br><span class="line">      <span class="keyword">if</span> <span class="keyword">not</span> self.y_dim:</span><br><span class="line">        h0 = lrelu(conv2d(image, self.df_dim, name=<span class="string">'d_h0_conv'</span>))</span><br><span class="line">        h1 = lrelu(self.d_bn1(conv2d(h0, self.df_dim*<span class="number">2</span>, name=<span class="string">'d_h1_conv'</span>)))</span><br><span class="line">        h2 = lrelu(self.d_bn2(conv2d(h1, self.df_dim*<span class="number">4</span>, name=<span class="string">'d_h2_conv'</span>)))</span><br><span class="line">        h3 = lrelu(self.d_bn3(conv2d(h2, self.df_dim*<span class="number">8</span>, name=<span class="string">'d_h3_conv'</span>)))</span><br><span class="line">        h4 = linear(tf.reshape(h3, [self.batch_size, <span class="number">-1</span>]), <span class="number">1</span>, <span class="string">'d_h4_lin'</span>)</span><br><span class="line"></span><br><span class="line">        <span class="keyword">return</span> tf.nn.sigmoid(h4), h4</span><br><span class="line">      <span class="keyword">else</span>:</span><br><span class="line">        yb = tf.reshape(y, [self.batch_size, <span class="number">1</span>, <span class="number">1</span>, self.y_dim])</span><br><span class="line">        x = conv_cond_concat(image, yb)</span><br><span class="line"></span><br><span class="line">        h0 = lrelu(conv2d(x, self.c_dim + self.y_dim, name=<span class="string">'d_h0_conv'</span>))</span><br><span class="line">        h0 = conv_cond_concat(h0, yb)</span><br><span class="line"></span><br><span class="line">        h1 = lrelu(self.d_bn1(conv2d(h0, self.df_dim + self.y_dim, name=<span class="string">'d_h1_conv'</span>)))</span><br><span class="line">        h1 = tf.reshape(h1, [self.batch_size, <span class="number">-1</span>])      </span><br><span class="line">        h1 = concat([h1, y], <span class="number">1</span>)</span><br><span class="line">        </span><br><span class="line">        h2 = lrelu(self.d_bn2(linear(h1, self.dfc_dim, <span class="string">'d_h2_lin'</span>)))</span><br><span class="line">        h2 = concat([h2, y], <span class="number">1</span>)</span><br><span class="line"></span><br><span class="line">        h3 = linear(h2, <span class="number">1</span>, <span class="string">'d_h3_lin'</span>)</span><br><span class="line">        </span><br><span class="line">        <span class="keyword">return</span> tf.nn.sigmoid(h3), h3</span><br><span class="line"></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">generator</span><span class="params">(self, z, y=None)</span>:</span></span><br><span class="line">    <span class="keyword">with</span> tf.variable_scope(<span class="string">"generator"</span>) <span class="keyword">as</span> scope:</span><br><span class="line">      <span class="keyword">if</span> <span class="keyword">not</span> self.y_dim:</span><br><span class="line">        s_h, s_w = self.output_height, self.output_width</span><br><span class="line">        <span class="comment"># 2 is stride</span></span><br><span class="line">        s_h2, s_w2 = conv_out_size_same(s_h, <span class="number">2</span>), conv_out_size_same(s_w, <span class="number">2</span>)</span><br><span class="line">        s_h4, s_w4 = conv_out_size_same(s_h2, <span class="number">2</span>), conv_out_size_same(s_w2, <span class="number">2</span>)</span><br><span class="line">        s_h8, s_w8 = conv_out_size_same(s_h4, <span class="number">2</span>), conv_out_size_same(s_w4, <span class="number">2</span>)</span><br><span class="line">        s_h16, s_w16 = conv_out_size_same(s_h8, <span class="number">2</span>), conv_out_size_same(s_w8, <span class="number">2</span>)</span><br><span class="line"></span><br><span class="line">
201        <span class="comment"># project `z` and reshape</span></span><br><span class="line">        self.z_, self.h0_w, self.h0_b = linear(</span><br><span class="line">            z, self.gf_dim*<span class="number">8</span>*s_h16*s_w16, <span class="string">'g_h0_lin'</span>, with_w=<span class="keyword">True</span>)</span><br><span class="line"></span><br><span class="line">        self.h0 = tf.reshape(</span><br><span class="line">            self.z_, [<span class="number">-1</span>, s_h16, s_w16, self.gf_dim * <span class="number">8</span>])</span><br><span class="line">        h0 = tf.nn.relu(self.g_bn0(self.h0))</span><br><span class="line"></span><br><span class="line">        self.h1, self.h1_w, self.h1_b = deconv2d(</span><br><span class="line">            h0, [self.batch_size, s_h8, s_w8, self.gf_dim*<span class="number">4</span>], name=<span class="string">'g_h1'</span>, with_w=<span class="keyword">True</span>)</span><br><span class="line">        h1 = tf.nn.relu(self.g_bn1(self.h1))</span><br><span class="line"></span><br><span class="line">        h2, self.h2_w, self.h2_b = deconv2d(</span><br><span class="line">            h1, [self.batch_size, s_h4, s_w4, self.gf_dim*<span class="number">2</span>], name=<span class="string">'g_h2'</span>, with_w=<span class="keyword">True</span>)</span><br><span class="line">        h2 = tf.nn.relu(self.g_bn2(h2))</span><br><span class="line"></span><br><span class="line">        h3, self.h3_w, self.h3_b = deconv2d(</span><br><span class="line">            h2, [self.batch_size, s_h2, s_w2, self.gf_dim*<span class="number">1</span>], name=<span class="string">'g_h3'</span>, with_w=<span class="keyword">True</span>)</span><br><span class="line">        h3 = tf.nn.relu(self.g_bn3(h3))</span><br><span class="line"></span><br><span class="line">        h4, self.h4_w, self.h4_b = deconv2d(</span><br><span class="line">            h3, [self.batch_size, s_h, s_w, self.c_dim], name=<span class="string">'g_h4'</span>, with_w=<span class="keyword">True</span>)</span><br><span class="line"></span><br><span class="line">        <span class="keyword">return</span> tf.nn.tanh(h4)</span><br><span class="line">      <span class="keyword">else</span>:</span><br><span class="line">        s_h, s_w = self.output_height, self.output_width</span><br><span class="line">        s_h2, s_h4 = int(s_h/<span class="number">2</span>), int(s_h/<span class="number">4</span>)</span><br><span class="line">        s_w2, s_w4 = int(s_w/<span class="number">2</span>), int(s_w/<span class="number">4</span>)</span><br><span class="line"></span><br><span class="line">        <span class="comment"># yb = tf.expand_dims(tf.expand_dims(y, 1),2)</span></span><br><span class="line">        yb = tf.reshape(y, [self.batch_size, <span class="number">1</span>, <span class="number">1</span>, self.y_dim])</span><br><span class="line">        z = concat([z, y], <span class="number">1</span>)</span><br><span class="line"></span><br><span class="line">        h0 = tf.nn.relu(</span><br><span class="line">            self.g_bn0(linear(z, self.gfc_dim, <span class="string">'g_h0_lin'</span>)))</span><br><span class="line">        h0 = concat([h0, y], <span class="number">1</span>)</span><br><span class="line"></span><br><span class="line">        h1 = tf.nn.relu(self.g_bn1(</span><br><span class="line">            linear(h0, self.gf_dim*<span class="number">2</span>*s_h4*s_w4, <span class="string">'g_h1_lin'</span>)))</span><br><span class="line">        h1 = tf.reshape(h1, [self.batch_size, s_h4, s_w4, self.gf_dim * <span class="number">2</span>])</span><br><span class="line"></span><br><span class="line">        h1 = conv_cond_concat(h1, yb)</span><br><span class="line"></span><br><span class="line">        h2 = tf.nn.relu(self.g_bn2(deconv2d(h1,</span><br><span class="line">            [self.batch_size, s_h2, s_w2, self.gf_dim * <span class="number">2</span>], name=<span class="string">'g_h2'</span>)))</span><br><span class="line">        h2 = conv_cond_concat(h2, yb)</span><br><span class="line"></span><br><span class="line">
201        <span class="keyword">return</span> tf.nn.sigmoid(</span><br><span class="line">            deconv2d(h2, [self.batch_size, s_h, s_w, self.c_dim], name=<span class="string">'g_h3'</span>))</span><br><span class="line"></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">sampler</span><span class="params">(self, z, y=None)</span>:</span> <span class="comment"># 采样测试</span></span><br><span class="line">    <span class="keyword">with</span> tf.variable_scope(<span class="string">"generator"</span>) <span class="keyword">as</span> scope:</span><br><span class="line">      scope.reuse_variables()</span><br><span class="line"></span><br><span class="line">      <span class="keyword">if</span> <span class="keyword">not</span> self.y_dim: <span class="comment"># generator</span></span><br><span class="line">        s_h, s_w = self.output_height, self.output_width</span><br><span class="line">        s_h2, s_w2 = conv_out_size_same(s_h, <span class="number">2</span>), conv_out_size_same(s_w, <span class="number">2</span>)</span><br><span class="line">        s_h4, s_w4 = conv_out_size_same(s_h2, <span class="number">2</span>), conv_out_size_same(s_w2, <span class="number">2</span>)</span><br><span class="line">        s_h8, s_w8 = conv_out_size_same(s_h4, <span class="number">2</span>), conv_out_size_same(s_w4, <span class="number">2</span>)</span><br><span class="line">        s_h16, s_w16 = conv_out_size_same(s_h8, <span class="number">2</span>), conv_out_size_same(s_w8, <span class="number">2</span>)</span><br><span class="line"></span><br><span class="line">        <span class="comment"># project `z` and reshape</span></span><br><span class="line">        h0 = tf.reshape(</span><br><span class="line">            linear(z, self.gf_dim*<span class="number">8</span>*s_h16*s_w16, <span class="string">'g_h0_lin'</span>),</span><br><span class="line">            [<span class="number">-1</span>, s_h16, s_w16, self.gf_dim * <span class="number">8</span>])</span><br><span class="line">        h0 = tf.nn.relu(self.g_bn0(h0, train=<span class="keyword">False</span>))</span><br><span class="line"></span><br><span class="line">        h1 = deconv2d(h0, [self.batch_size, s_h8, s_w8, self.gf_dim*<span class="number">4</span>], name=<span class="string">'g_h1'</span>)</span><br><span class="line">        h1 = tf.nn.relu(self.g_bn1(h1, train=<span class="keyword">False</span>))</span><br><span class="line"></span><br><span class="line">        h2 = deconv2d(h1, [self.batch_size, s_h4, s_w4, self.gf_dim*<span class="number">2</span>], name=<span class="string">'g_h2'</span>)</span><br><span class="line">        h2 = tf.nn.relu(self.g_bn2(h2, train=<span class="keyword">False</span>))</span><br><span class="line"></span><br><span class="line">        h3 = deconv2d(h2, [self.batch_size, s_h2, s_w2, self.gf_dim*<span class="number">1</span>], name=<span class="string">'g_h3'</span>)</span><br><span class="line">        h3 = tf.nn.relu(self.g_bn3(h3, train=<span class="keyword">False</span>))</span><br><span class="line"></span><br><span class="line">        h4 = deconv2d(h3, [self.batch_size, s_h, s_w, self.c_dim], name=<span class="string">'g_h4'</span>)</span><br><span class="line"></span><br><span class="line">        <span class="keyword">return</span> tf.nn.tanh(h4)</span><br><span class="line">      <span class="keyword">else</span>: <span class="comment"># discriminator</span></span><br><span class="line">        s_h, s_w = self.output_height, self.output_width</span><br><span class="line">        s_h2, s_h4 = int(s_h/<span class="number">2</span>), int(s_h/<span class="number">4</span>)</span><br><span class="line">        s_w2, s_w4 = int(s_w/<span class="number">2</span>), int(s_w/<span class="number">4</span>)</span><br><span class="line"></span><br><span class="line">        <span class="comment"># yb = tf.reshape(y, [-1, 1, 1, self.y_dim])</span></span><br><span class="line">        yb = tf.reshape(y, [self.batch_size, <span class="number">1</span>, <span class="number">1</span>, self.y_dim])</span><br><span class="line">        z = concat([z, y], <span class="number">1</span>)</span><br><span class="line"></span><br><span class="line">        h0 = tf.nn.relu(self.g_bn0(linear(z, self.gfc_dim, <span class="string">'g_h0_lin'</span>), train=<span class="keyword">False</span>))</span><br><span class="line">        h0 = concat([h0, y], <span class="number">1</span>)</span><br><span class="line"></span><br><span class="line">        h1 = tf.nn.relu(self.g_bn1(</span><br><span class="line">            linear(h0, self.gf_dim*<span class="number">2</span>*s_h4*s_w4, <span class="string">'g_h1_lin'</span>), train=<span class="keyword">False</span>))</span><br><span class="line">        h1 = tf.reshape(h1, [self.batch_size, s_h4, s_w4, self.gf_dim * <span class="number">2</span>])</span><br><span class="line">        h1 = conv_cond_concat(h1, yb)</span><br><span class="line"></span><br><span class="line">        h2 = tf.nn.relu(self.g_bn2(</span><br><span class="line">            deconv2d(h1, [self.batch_size, s_h2, s_w2, self.gf_dim * <span class="number">2</span>], name=<span class="string">
201'g_h2'</span>), train=<span class="keyword">False</span>))</span><br><span class="line">        h2 = conv_cond_concat(h2, yb)</span><br><span class="line"></span><br><span class="line">        <span class="keyword">return</span> tf.nn.sigmoid(deconv2d(h2, [self.batch_size, s_h, s_w, self.c_dim], name=<span class="string">'g_h3'</span>))</span><br><span class="line"></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">load_mnist</span><span class="params">(self)</span>:</span></span><br><span class="line">    data_dir = os.path.join(self.data_dir, self.dataset_name)</span><br><span class="line">    </span><br><span class="line">    fd = open(os.path.join(data_dir,<span class="string">'train-images-idx3-ubyte'</span>))</span><br><span class="line">    loaded = np.fromfile(file=fd,dtype=np.uint8)</span><br><span class="line">    trX = loaded[<span class="number">16</span>:].reshape((<span class="number">60000</span>,<span class="number">28</span>,<span class="number">28</span>,<span class="number">1</span>)).astype(np.float)</span><br><span class="line"></span><br><span class="line">    fd = open(os.path.join(data_dir,<span class="string">'train-labels-idx1-ubyte'</span>))</span><br><span class="line">    loaded = np.fromfile(file=fd,dtype=np.uint8)</span><br><span class="line">    trY = loaded[<span class="number">8</span>:].reshape((<span class="number">60000</span>)).astype(np.float)</span><br><span class="line"></span><br><span class="line">    fd = open(os.path.join(data_dir,<span class="string">'t10k-images-idx3-ubyte'</span>))</span><br><span class="line">    loaded = np.fromfile(file=fd,dtype=np.uint8)</span><br><span class="line">    teX = loaded[<span class="number">16</span>:].reshape((<span class="number">10000</span>,<span class="number">28</span>,<span class="number">28</span>,<span class="number">1</span>)).astype(np.float)</span><br><span class="line"></span><br><span class="line">    fd = open(os.path.join(data_dir,<span class="string">'t10k-labels-idx1-ubyte'</span>))</span><br><span class="line">    loaded = np.fromfile(file=fd,dtype=np.uint8)</span><br><span class="line">    teY = loaded[<span class="number">8</span>:].reshape((<span class="number">10000</span>)).astype(np.float)</span><br><span class="line"></span><br><span class="line">    trY = np.asarray(trY)</span><br><span class="line">    teY = np.asarray(teY)</span><br><span class="line">    </span><br><span class="line">    X = np.concatenate((trX, teX), axis=<span class="number">0</span>)</span><br><span class="line">    y = np.concatenate((trY, teY), axis=<span class="number">0</span>).astype(np.int)</span><br><span class="line">    </span><br><span class="line">    seed = <span class="number">547</span></span><br><span class="line">    np.random.seed(seed)</span><br><span class="line">    np.random.shuffle(X)</span><br><span class="line">    np.random.seed(seed)</span><br><span class="line">    np.random.shuffle(y)</span><br><span class="line">    </span><br><span class="line">    y_vec = np.zeros((len(y), self.y_dim), dtype=np.float)</span><br><span class="line">    <span class="keyword">for</span> i, label <span class="keyword">in</span> enumerate(y):</span><br><span class="line">      y_vec[i,y[i]] = <span class="number">1.0</span></span><br><span class="line">    </span><br><span class="line">    <span class="keyword">return</span> X/<span class="number">255.</span>,y_vec</span><br><span class="line"></span><br><span class="line"><span class="meta">  @property # 可以当属性来用</span></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">model_dir</span><span class="params">(self)</span>:</span></span><br><span class="line">    <span class="keyword">return</span> <span class="string">"&#123;&#125;_&#123;&#125;_&#123;&#125;_&#123;&#125;"</span>.format(</span><br><span class="line">        self.dataset_name, self.batch_size,</span><br><span class="line">        self.output_height, self.output_width)</span><br><span class="line">      </span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">save</span><span class="params">(self, checkpoint_dir, step)</span>:</span></span><br><span class="line">    <span class="comment"># save checkpoint files</span></span><br><span class="line">    model_name = <span class="string">"DCGAN.model"</span></span><br><span class="line">    checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir)</span><br><span class="line"></span><br><span class="line">    <span class="keyword">if</span> <span class="keyword">not</span> os.path.exists(checkpoint_dir):</span><br><span class="line">      os.makedirs(checkpoint_dir)</span><br><span class="line"></span><br><span class="line">    self.saver.save(self.sess,</span><br><span class="line">            os.path.join(checkpoint_dir, model_name),</span><br><span class="line">            global_step=step)</span><br><span class="line"></span><br><span class="line">  <span class="comment"># load checkpoints file</span></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">load</span><span class="params">(self, checkpoint_dir)</span>:</span></span><br><span class="line">    <span class="keyword">
201import</span> re</span><br><span class="line">    print(<span class="string">" [*] Reading checkpoints..."</span>)</span><br><span class="line">    checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir)</span><br><span class="line">    <span class="comment">#A CheckpointState if the state was available, None</span></span><br><span class="line">    <span class="comment"># otherwise</span></span><br><span class="line">    ckpt = tf.train.get_checkpoint_state(checkpoint_dir)</span><br><span class="line">    <span class="keyword">if</span> ckpt <span class="keyword">and</span> ckpt.model_checkpoint_path:</span><br><span class="line">      <span class="comment"># basename:Returns the final component of a pathname</span></span><br><span class="line">      ckpt_name = os.path.basename(ckpt.model_checkpoint_path)</span><br><span class="line">      self.saver.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))</span><br><span class="line">      counter = int(next(re.finditer(<span class="string">"(\d+)(?!.*\d)"</span>,ckpt_name)).group(<span class="number">0</span>))</span><br><span class="line">      print(<span class="string">" [*] Success to read &#123;&#125;"</span>.format(ckpt_name))</span><br><span class="line">      <span class="keyword">return</span> <span class="keyword">True</span>, counter</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      print(<span class="string">" [*] Failed to find a checkpoint"</span>)</span><br><span class="line">      <span class="keyword">return</span> <span class="keyword">False</span>, <span class="number">0</span></span><br></pre></td></tr></table></figure>
202<ul>
203<li><span class="lang:python decode:true      crayon-inline ">from __future__ import division</span> 这句话当python的版本为2.x时生效,可以让两个整数数字相除的结果返回一个浮点数(在python2中默认是整数,python3默认为浮点数)。glob可以以简单的正则表达式筛选的方式返回某个文件夹下符合要求的文件名列表。</li>
204<li>DCGAN的构造方法除了设置一大堆的属性之外,还要注意区分dataset是否是mnist,因为mnist是灰度图像,所以应该设置channel = 1( <span class="lang:default decode:true      crayon-inline ">self.c_dim = 1</span> ),如果是彩色图像,则 <span class="lang:default decode:true      crayon-inline ">self.c_dim = 3</span> or  <span class="lang:default decode:true      crayon-inline ">self.c_dim = 4</span> 。然后就是<strong>build_model</strong>。</li>
205<li><span class="lang:default decode:true      crayon-inline ">self.generator</span> 用于构造生成器; <span class="lang:default decode:true      crayon-inline ">self.discriminator</span> 用于构造鉴别器; <span class="lang:default decode:true      crayon-inline ">self.sampler</span> 用于随机采样(用于生成样本)。这里需要注意的是, <span class="lang:default decode:true      crayon-inline ">self.y</span> 只有当dataset是mnist的时候才不为None,不是mnist的情况下,只需要 <span class="lang:default decode:true      crayon-inline ">self.z</span> 即可生成samples。</li>
206<li><span class="lang:default decode:true      crayon-inline ">sigmoid_cross_entropy_with_logits</span> 函数被重新定义了,是为了兼容不同版本的tensorflow。该函数首先使用sigmoid activation,然后计算cross-entropy loss。</li>
207<li><span class="lang:default decode:true      crayon-inline ">self.g_loss</span> 是生成器损失; <span class="lang:default decode:true      crayon-inline ">self.d_loss_real</span> 是真实图片的鉴别器损失; <span class="lang:default decode:true      crayon-inline ">self.d_loss_fake</span> 是虚假图片(由生成器生成的fake images)的损失; <span class="lang:default decode:true      crayon-inline ">self.d_loss</span> 是总的鉴别器损失。</li>
208<li>这里的 <span class="lang:default decode:true      crayon-inline ">histogram_summary</span> 和 <span class="lang:default decode:true      crayon-inline ">scalar_summary</span> 是为了在后续在tensorboard中对各个损失函数进行可视化。</li>
209<li><span class="lang:default decode:true      crayon-inline ">tf.trainable_variables()</span> 可以获取model的全部可训练参数,由于我们在定义生成器和鉴别器变量的时候使用了不同的name,因此我们可以通过variable的name来获取得到<strong>self.d_vars</strong>(鉴别器相关变量),<strong>self.g_vars</strong>(生成器相关变量)。 <span class="lang:default decode:true      crayon-inline ">self.saver = tf.train.Saver()</span> 用于保存训练好的模型参数到checkpoint。</li>
210<li><span class="lang:default decode:true      crayon-inline ">train</span> 函数是核心的训练函数。这里optimizer和DCGAN的原文保持一直,选用<strong>Adam</strong>优化函数, <span class="lang:default decode:true      crayon-inline ">lr=0.0002</span> , <span class="lang:default decode:true      crayon-inline ">beta1=0.5</span> 。 <span class="lang:default decode:true      crayon-inline ">merge_summary</span> 函数和 <span class="lang:default decode:true      crayon-inline ">SummaryWriter</span> 用于构建summary,在tensorboard中显示。</li>
211<li><span class="lang:default decode:true      crayon-inline ">sample_z</span> 是从[-1,1]的均匀分布产生的。如果dataset是mnist,则可以直接读取<strong>sample_inputs</strong>和<strong>sample_labels</strong>
211。否则需要手动逐个处理图像, <span class="lang:default decode:true      crayon-inline ">get_image</span><br>返回的是取值为(-1,1)的,shape为(resize_height,resize_width)的ndarray。如果处理的图像是灰度图像,则需要再增加一个dim,表示图像的<strong>channel=1</strong>,对应的代码是 <span class="lang:default decode:true      crayon-inline ">sample_inputs = np.array(sample).astype(np.float32)[:, :, :, None]</span> 。</li>
212<li>接下来通过 <span class="lang:default decode:true      crayon-inline ">self.sess.run([d_optim,…</span> 和 <span class="lang:default decode:true      crayon-inline ">self.sess.run([g_optim,…)</span> 来更新鉴别器和生成器。 <span class="lang:default decode:true      crayon-inline ">self.writer.add_summary(summary_str, counter)</span> 增加summary到writer。由于同样的原因,这里仍然需要区分mnist和其他的数据集,所以计算最优化函数的过程需要一个<strong>if</strong>和一个<strong>else</strong>。</li>
213<li><span class="lang:default decode:true      crayon-inline ">np.mod(counter, config.print_every) == 1</span> 表示每<em>print_every</em>次生成一次samples; <span class="lang:default decode:true      crayon-inline ">np.mod(counter, config.checkpoint_every) == 2</span> 表示每<em>checkpoint_every</em>次保存一下checkpoint file。</li>
214<li>下面是discriminator(鉴别器)的具体实现。首先鉴别器使用<strong>conv</strong>(卷积)操作,激活函数使用<strong>leaky-relu</strong>,每一个layer需要使用batch normalization。tensorflow的batch normalization使用 <span class="lang:default decode:true      crayon-inline ">tf.contrib.layers.batch_norm</span> 实现。如果不是mnist,则第一层使用<strong>leaky-relu+conv2d</strong>,后面三层都使用<strong>conv2d+BN+leaky-relu</strong>,最后加上一个one hidden unit的linear layer,再送入sigmoid函数即可;如果是mnist,则 <span class="lang:default decode:true      crayon-inline ">yb = tf.reshape(y, [self.batch_size, 1, 1, self.y_dim])</span> 首先给y增加两维,以便可以和image连接起来,这里实际上是使用了conditional GAN(条件GAN)的思想。 <span class="lang:default decode:true      crayon-inline ">x = conv_cond_concat(image, yb)</span> 得到condition和image合并之后的结果,然后 <span class="lang:default decode:true      crayon-inline ">h0 = lrelu(conv2d(x, self.c_dim + self.y_dim, name=’d_h0_conv’))</span> 进行卷积操作。第二次进行<strong>conv2d+leaky-relu+concat</strong>操作。第三次进行<strong>conv2d+BN+leaky-relu+reshape+concat</strong>操作。第四次进行<strong>linear+BN+leaky-relu+concat</strong>操作。最后同样是<strong>linear+sigmoid</strong>操作。</li>
215<li>下面是generator(生成器)的具体实现。和discriminator不同的是,generator需要使用deconv(反卷积)以及relu 激活函数。generator的结构是:1.如果不是mnist:<strong>linear+reshape+BN+relu—-&gt;(deconv+BN+relu)x3 —-&gt;deconv+tanh</strong>;2.如果是mnist,则除了需要考虑输入z之外,还需要考虑label y,即需要将z和y连接起来(Conditional GAN),具体的结构是:reshape+concat—-&gt;linear+BN+relu+concat—-&gt;linear+BN+relu+reshape+concat—-&gt;deconv+BN+relu+concat—-&gt;deconv+sigmoid。注意的最后的激活函数没有采用通常的tanh,而是采用了sigmoid(其输出会直接映射到0-1之间)。</li>
216<li>sampler函数是采样函数,用于生成样本送入当前训练的生成器,查看训练效果。其逻辑和generator函数基本类似,也是需要区分是否是mnist,二者需要采用不同的结构。不是mnist时,y=None即可;否则mnist还需要考虑y。</li>
217<li><span class="lang:default decode:true  crayon-inline ">load_mnist</span> 函数用于加载mnist数据集; <span class="lang:default decode:true  crayon-inline ">save</span> 函数用于保存checkpoint; <span class="lang:default decode:true  crayon-inline ">load</span> 函数用于加载checkpoint。</li>
218</ul>
219<h4 id="2-4-ops-py"><a href="#2-4-ops-py" class="headerlink" title="2.4 ops.py"></a>2.4 <strong>ops.py</strong></h4><h4 id="nbsp-nbsp-nbsp-nbsp-ops-py代码如下"><a href="#nbsp-nbsp-nbsp-nbsp-ops-py代码如下" class="headerlink" title="&nbsp;&nbsp;&nbsp;&nbsp;ops.py代码如下:"></a>&nbsp;&nbsp;&nbsp;&nbsp;ops.py代码如下:</h4><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br><span class="line">44</span><br><span class="line">45</span><br><span class="line">46</span><br><span class="line">47</span><br><span class="line">48</span><br><span class="line">49</span><br><span class="line">50</span><br><span class="line">51</span><br><span class="line">52</span><br><span class="line">
21953</span><br><span class="line">54</span><br><span class="line">55</span><br><span class="line">56</span><br><span class="line">57</span><br><span class="line">58</span><br><span class="line">59</span><br><span class="line">60</span><br><span class="line">61</span><br><span class="line">62</span><br><span class="line">63</span><br><span class="line">64</span><br><span class="line">65</span><br><span class="line">66</span><br><span class="line">67</span><br><span class="line">68</span><br><span class="line">69</span><br><span class="line">70</span><br><span class="line">71</span><br><span class="line">72</span><br><span class="line">73</span><br><span class="line">74</span><br><span class="line">75</span><br><span class="line">76</span><br><span class="line">77</span><br><span class="line">78</span><br><span class="line">79</span><br><span class="line">80</span><br><span class="line">81</span><br><span class="line">82</span><br><span class="line">83</span><br><span class="line">84</span><br><span class="line">85</span><br><span class="line">86</span><br><span class="line">87</span><br><span class="line">88</span><br><span class="line">89</span><br><span class="line">90</span><br><span class="line">91</span><br><span class="line">92</span><br><span class="line">93</span><br><span class="line">94</span><br><span class="line">95</span><br><span class="line">96</span><br><span class="line">97</span><br><span class="line">98</span><br><span class="line">99</span><br><span class="line">100</span><br><span class="line">101</span><br><span class="line">102</span><br><span class="line">103</span><br><span class="line">104</span><br><span class="line">105</span><br><span class="line">106</span><br><span class="line">107</span><br><span class="line">108</span><br><span class="line">109</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">import</span> math</span><br><span class="line"><span class="keyword">import</span> numpy <span class="keyword">as</span> np </span><br><span class="line"><span class="keyword">import</span> tensorflow <span class="keyword">as</span> tf</span><br><span class="line"></span><br><span class="line"><span class="keyword">from</span> tensorflow.python.framework <span class="keyword">import</span> ops</span><br><span class="line"></span><br><span class="line"><span class="keyword">from</span> utils <span class="keyword">import</span> *</span><br><span class="line"></span><br><span class="line"><span class="keyword">try</span>:</span><br><span class="line">  image_summary = tf.image_summary</span><br><span class="line">  scalar_summary = tf.scalar_summary</span><br><span class="line">  histogram_summary = tf.histogram_summary</span><br><span class="line">  merge_summary = tf.merge_summary</span><br><span class="line">  SummaryWriter = tf.train.SummaryWriter</span><br><span class="line"><span class="keyword">except</span>:</span><br><span class="line">  image_summary = tf.summary.image</span><br><span class="line">  scalar_summary = tf.summary.scalar</span><br><span class="line">  histogram_summary = tf.summary.histogram</span><br><span class="line">  merge_summary = tf.summary.merge</span><br><span class="line">  SummaryWriter = tf.summary.FileWriter</span><br><span class="line"></span><br><span class="line"><span class="keyword">if</span> <span class="string">"concat_v2"</span> <span class="keyword">in</span> dir(tf):</span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">concat</span><span class="params">(tensors, axis, *args, **kwargs)</span>:</span></span><br><span class="line">    <span class="keyword">return</span> tf.concat_v2(tensors, axis, *args, **kwargs)</span><br><span class="line"><span class="keyword">else</span>:</span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">concat</span><span class="params">(tensors, axis, *args, **kwargs)</span>:</span></span><br><span class="line">    <span class="keyword">return</span> tf.concat(tensors, axis, *args, **kwargs)</span><br><span class="line"></span><br><span class="line"><span class="class"><span class="keyword">class</span> <span class="title">batch_norm</span><span class="params">(object)</span>:</span></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">__init__</span><span class="params">(self, epsilon=<span class="number">1e-5</span>, momentum = <span class="number">
2190.9</span>, name=<span class="string">"batch_norm"</span>)</span>:</span></span><br><span class="line">    <span class="keyword">with</span> tf.variable_scope(name):</span><br><span class="line">      self.epsilon  = epsilon</span><br><span class="line">      self.momentum = momentum</span><br><span class="line">      self.name = name</span><br><span class="line"></span><br><span class="line">  <span class="comment"># 定义了class 的__call__ 方法,可以把类像函数一样调用</span></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">__call__</span><span class="params">(self, x, train=True)</span>:</span></span><br><span class="line">    <span class="keyword">return</span> tf.contrib.layers.batch_norm(x,</span><br><span class="line">                      decay=self.momentum, </span><br><span class="line">                      updates_collections=<span class="keyword">None</span>,</span><br><span class="line">                      epsilon=self.epsilon,</span><br><span class="line">                      scale=<span class="keyword">True</span>,</span><br><span class="line">                      is_training=train,</span><br><span class="line">                      scope=self.name)</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">conv_cond_concat</span><span class="params">(x, y)</span>:</span></span><br><span class="line">  <span class="string">"""Concatenate conditioning vector on feature map axis."""</span></span><br><span class="line">  x_shapes = x.get_shape()</span><br><span class="line">  y_shapes = y.get_shape()</span><br><span class="line">  <span class="comment"># 沿axis = 3(最后一个维度连接)</span></span><br><span class="line">  <span class="keyword">return</span> concat([</span><br><span class="line">    x, y*tf.ones([x_shapes[<span class="number">0</span>], x_shapes[<span class="number">1</span>], x_shapes[<span class="number">2</span>], y_shapes[<span class="number">3</span>
219]])], <span class="number">3</span>)</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">conv2d</span><span class="params">(input_, output_dim, </span></span></span><br><span class="line"><span class="function"><span class="params">       k_h=<span class="number">5</span>, k_w=<span class="number">5</span>, d_h=<span class="number">2</span>, d_w=<span class="number">2</span>, stddev=<span class="number">0.02</span>,</span></span></span><br><span class="line"><span class="function"><span class="params">       name=<span class="string">"conv2d"</span>)</span>:</span></span><br><span class="line">  <span class="keyword">with</span> tf.variable_scope(name):</span><br><span class="line">    w = tf.get_variable(<span class="string">'w'</span>, [k_h, k_w, input_.get_shape()[<span class="number">-1</span>], output_dim],</span><br><span class="line">              initializer=tf.truncated_normal_initializer(stddev=stddev))</span><br><span class="line">    conv = tf.nn.conv2d(input_, w, strides=[<span class="number">1</span>, d_h, d_w, <span class="number">1</span>], padding=<span class="string">'SAME'</span>)</span><br><span class="line"></span><br><span class="line">    biases = tf.get_variable(<span class="string">'biases'</span>, [output_dim], initializer=tf.constant_initializer(<span class="number">0.0</span>))</span><br><span class="line">    conv = tf.reshape(tf.nn.bias_add(conv, biases), conv.get_shape())</span><br><span class="line"></span><br><span class="line">    <span class="keyword">return</span> conv</span><br><span class="line"></span><br><span class="line"><span class="comment"># 做一个反卷积操作,tf.nn.conv2d_transpose</span></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">deconv2d</span><span class="params">(input_, output_shape,</span></span></span><br><span class="line"><span class="function"><span class="params">       k_h=<span class="number">5</span>, k_w=<span class="number">5</span>, d_h=<span class="number">2</span>, d_w=<span class="number">2</span>, stddev=<span class="number">0.02</span>,</span></span></span><br><span class="line"><span class="function"><span class="params">       name=<span class="string">"deconv2d"</span>, with_w=False)</span>:</span></span><br><span class="line">  <span class="keyword">with</span> tf.variable_scope(name):</span><br><span class="line">    <span class="comment"># filter : [height, width, output_channels, in_channels]</span></span><br><span class="line">    w = tf.get_variable(<span class="string">'w'</span>, [k_h, k_w, output_shape[<span class="number">-1</span>], input_.get_shape()[<span class="number">-1</span>]],</span><br><span class="line">              initializer=tf.random_normal_initializer(stddev=stddev))</span><br><span class="line">    </span><br><span class="line">    <span class="keyword">try</span>:</span><br><span class="line">      deconv = tf.nn.conv2d_transpose(input_, w, output_shape=output_shape,</span><br><span class="line">                strides=[<span class="number">1</span>, d_h, d_w, <span class="number">1</span>])</span><br><span class="line"></span><br><span class="line">    <span class="comment"># Support for verisons of TensorFlow before 0.7.0</span></span><br><span class="line">    <span class="keyword">except</span> AttributeError:</span><br><span class="line">      deconv = tf.nn.deconv2d(input_, w, output_shape=output_shape,</span><br><span class="line">                strides=[<span class="number">1</span>, d_h, d_w, <span class="number">1</span>])</span><br><span class="line"></span><br><span class="line">    biases = tf.get_variable(<span class="string">'biases'</span>, [output_shape[<span class="number">-1</span>]], initializer=tf.constant_initializer(<span class="number">0.0</span>))</span><br><span class="line">    deconv = tf.reshape(tf.nn.bias_add(deconv, biases), deconv.get_shape())</span><br><span class="line"></span><br><span class="line">    <span class="keyword">if</span> with_w:</span><br><span class="line">
219      <span class="keyword">return</span> deconv, w, biases</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      <span class="keyword">return</span> deconv</span><br><span class="line"></span><br><span class="line"><span class="comment"># leaky relu</span></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">lrelu</span><span class="params">(x, leak=<span class="number">0.2</span>, name=<span class="string">"lrelu"</span>)</span>:</span></span><br><span class="line">  <span class="keyword">return</span> tf.maximum(x, leak*x)</span><br><span class="line"></span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">linear</span><span class="params">(input_, output_size, scope=None, stddev=<span class="number">0.02</span>, bias_start=<span class="number">0.0</span>, with_w=False)</span>:</span></span><br><span class="line">  <span class="comment"># 本质其实就是做了一个matmul....</span></span><br><span class="line">  shape = input_.get_shape().as_list()</span><br><span class="line"></span><br><span class="line">  <span class="keyword">with</span> tf.variable_scope(scope <span class="keyword">or</span> <span class="string">"Linear"</span>):</span><br><span class="line">    matrix = tf.get_variable(<span class="string">"Matrix"</span>, [shape[<span class="number">1</span>], output_size], tf.float32,</span><br><span class="line">                 tf.random_normal_initializer(stddev=stddev))</span><br><span class="line">    bias = tf.get_variable(<span class="string">"bias"</span>, [output_size],</span><br><span class="line">      initializer=tf.constant_initializer(bias_start))</span><br><span class="line">    <span class="keyword">if</span> with_w:</span><br><span class="line">      <span class="keyword">return</span>
219 tf.matmul(input_, matrix) + bias, matrix, bias</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      <span class="keyword">return</span> tf.matmul(input_, matrix) + bias</span><br></pre></td></tr></table></figure>
220<ul>
221<li>第9行到第20行的代码是为了保持tf0.x和tf1.x版本的兼容性。tf0.x版本使用<strong>tf.xxx_summary</strong>风格的函数,而tf1.x版本则使用tf.summary.xxx风格的函数。为了保持一致性,通过重命名统一成<strong>tf.xxx_summary</strong>风格了。</li>
222<li>22行到27行重新定义了concat函数,也是为了兼容性考虑, <span class="lang:default decode:true    crayon-inline ">if “concat_v2” in dir(tf):</span> 这句话是说如果tf有concat_v2这个方法的话,tf0.x中使用<strong>concat_v2</strong>函数,而tf1.x版本中使用<strong>concat</strong>函数。</li>
223<li>29行到44行定义了<strong>batch_norm</strong>类。需要注意的是37-44行定义了类的__call__特殊方法,这个方法的作用是可以将类像普通的函数那样直接调用,而不用先构造一个对象再调用方法,这是常用的一个技巧。tf中的batch normalization 是函数 <span class="lang:default decode:true    crayon-inline ">tf.contrib.layers.batch_norm</span> </li>
224<li><strong>conv_cond_concat</strong>函数的作用是将conv(卷积)和cond(条件)concat起来。在mnist的generator和discriminator中会用到。</li>
225<li>54行到65行的<strong>conv2d</strong>函数重新定义了卷积操作,主要是封装了 <span class="lang:default decode:true    crayon-inline ">tf.nn.conv2d</span> 函数。</li>
226<li>68行到91行定义了<strong>deconv2d</strong>(反卷积)函数。tf0.x的反卷积函数为 <span class="lang:default decode:true    crayon-inline ">tf.nn.deconv2d</span> ,tf1.x的反卷积函数为 <span class="lang:default decode:true    crayon-inline ">tf.nn.conv2d_transpose</span> 。最后还加上了一个bias( <span class="lang:default decode:true    crayon-inline ">tf.nn.bias_add</span> )。</li>
227<li>94到95行定义了leaky-relu函数<strong>lrelu</strong>。其实就一行代码: <span class="lang:default decode:true    crayon-inline ">tf.maximum(x, leak*x)</span> 。</li>
228<li>97行到109行定义了linear函数,其实就是一个<strong>fully_connected layer</strong>。</li>
229</ul>
230<h4 id="2-5-utils-py"><a href="#2-5-utils-py" class="headerlink" title="2.5 utils.py"></a>2.5 <strong>utils.py</strong></h4><h4 id="nbsp-nbsp-nbsp-nbsp-utils-py代码如下"><a href="#nbsp-nbsp-nbsp-nbsp-utils-py代码如下" class="headerlink" title="&nbsp;&nbsp;&nbsp;&nbsp;utils.py代码如下:"></a>&nbsp;&nbsp;&nbsp;&nbsp;utils.py代码如下:</h4><figure class="highlight plain"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br><span class="line">44</span><br><span class="line">45</span><br><span class="line">46</span><br><span class="line">47</span><br><span class="line">48</span><br><span class="line">49</span><br><span class="line">50</span><br><span class="line">51</span><br><span class="line">52</span><br><span class="line">
23053</span><br><span class="line">54</span><br><span class="line">55</span><br><span class="line">56</span><br><span class="line">57</span><br><span class="line">58</span><br><span class="line">59</span><br><span class="line">60</span><br><span class="line">61</span><br><span class="line">62</span><br><span class="line">63</span><br><span class="line">64</span><br><span class="line">65</span><br><span class="line">66</span><br><span class="line">67</span><br><span class="line">68</span><br><span class="line">69</span><br><span class="line">70</span><br><span class="line">71</span><br><span class="line">72</span><br><span class="line">73</span><br><span class="line">74</span><br><span class="line">75</span><br><span class="line">76</span><br><span class="line">77</span><br><span class="line">78</span><br><span class="line">79</span><br><span class="line">80</span><br><span class="line">81</span><br><span class="line">82</span><br><span class="line">83</span><br><span class="line">84</span><br><span class="line">85</span><br><span class="line">86</span><br><span class="line">87</span><br><span class="line">88</span><br><span class="line">89</span><br><span class="line">90</span><br><span class="line">91</span><br><span class="line">92</span><br><span class="line">93</span><br><span class="line">94</span><br><span class="line">95</span><br><span class="line">96</span><br><span class="line">97</span><br><span class="line">98</span><br><span class="line">99</span><br><span class="line">100</span><br><span class="line">101</span><br><span class="line">102</span><br><span class="line">103</span><br><span class="line">104</span><br><span class="line">105</span><br><span class="line">106</span><br><span class="line">107</span><br><span class="line">108</span><br><span class="line">109</span><br><span class="line">110</span><br><span class="line">111</span><br><span class="line">112</span><br><span class="line">113</span><br><span class="line">114</span><br><span class="line">115</span><br><span class="line">116</span><br><span class="line">117</span><br><span class="line">118</span><br><span class="line">119</span><br><span class="line">120</span><br><span class="line">121</span><br><span class="line">122</span><br><span class="line">123</span><br><span class="line">124</span><br><span class="line">125</span><br><span class="line">126</span><br><span class="line">127</span><br><span class="line">128</span><br><span class="line">129</span><br><span class="line">130</span><br><span class="line">131</span><br><span class="line">132</span><br><span class="line">133</span><br><span class="line">134</span><br><span class="line">135</span><br><span class="line">136</span><br><span class="line">137</span><br><span class="line">138</span><br><span class="line">139</span><br><span class="line">140</span><br><span class="line">141</span><br><span class="line">142</span><br><span class="line">143</span><br><span class="line">144</span><br><span class="line">145</span><br><span class="line">146</span><br><span class="line">147</span><br><span class="line">148</span><br><span class="line">149</span><br><span class="line">150</span><br><span class="line">151</span><br><span class="line">152</span><br><span class="line">153</span><br><span class="line">154</span><br><span class="line">155</span><br><span class="line">156</span><br><span class="line">157</span><br><span class="line">158</span><br><span class="line">159</span><br><span class="line">160</span><br><span class="line">161</span><br><span class="line">162</span><br><span class="line">163</span><br><span class="line">164</span><br><span class="line">165</span><br><span class="line">166</span><br><span class="line">167</span><br><span class="line">168</span><br><span class="line">169</span><br><span class="line">170</span><br><span class="line">171</span><br><span class="line">172</span><br><span class="line">173</span><br><span class="line">174</span><br><span class="line">175</span><br><span class="line">176</span><br><span class="line">177</span><br><span class="line">178</span><br><span class="line">179</span><br><span class="line">180</span><br><span class="line">181</span><br><span class="line">182</span><br><span class="line">183</span><br><span class="line">184</span><br><span class="line">185</span><br><span class="line">186</span><br><span class="line">187</span><br><span class="line">188</span><br><span class="line">189</span><br><span class="line">190</span><br><span class="line">191</span><br><span class="line">192</span><br><span class="line">193</span><br><span class="line">
230194</span><br><span class="line">195</span><br><span class="line">196</span><br><span class="line">197</span><br><span class="line">198</span><br><span class="line">199</span><br><span class="line">200</span><br><span class="line">201</span><br><span class="line">202</span><br><span class="line">203</span><br><span class="line">204</span><br><span class="line">205</span><br><span class="line">206</span><br><span class="line">207</span><br><span class="line">208</span><br><span class="line">209</span><br><span class="line">210</span><br><span class="line">211</span><br><span class="line">212</span><br><span class="line">213</span><br><span class="line">214</span><br><span class="line">215</span><br><span class="line">216</span><br><span class="line">217</span><br><span class="line">218</span><br><span class="line">219</span><br><span class="line">220</span><br><span class="line">221</span><br><span class="line">222</span><br><span class="line">223</span><br><span class="line">224</span><br><span class="line">225</span><br><span class="line">226</span><br><span class="line">227</span><br><span class="line">228</span><br><span class="line">229</span><br><span class="line">230</span><br><span class="line">231</span><br><span class="line">232</span><br><span class="line">233</span><br><span class="line">234</span><br><span class="line">235</span><br><span class="line">236</span><br><span class="line">237</span><br><span class="line">238</span><br><span class="line">239</span><br><span class="line">240</span><br><span class="line">241</span><br><span class="line">242</span><br><span class="line">243</span><br><span class="line">244</span><br><span class="line">245</span><br><span class="line">246</span><br><span class="line">247</span><br><span class="line">248</span><br><span class="line">249</span><br><span class="line">250</span><br><span class="line">251</span><br><span class="line">252</span><br><span class="line">253</span><br><span class="line">254</span><br><span class="line">255</span><br><span class="line">256</span><br><span class="line">257</span><br><span class="line">258</span><br><span class="line">259</span><br><span class="line">260</span><br><span class="line">261</span><br><span class="line">262</span><br><span class="line">263</span><br><span class="line">264</span><br><span class="line">265</span><br><span class="line">266</span><br><span class="line">267</span><br><span class="line">268</span><br><span class="line">269</span><br><span class="line">270</span><br><span class="line">271</span><br><span class="line">272</span><br><span class="line">273</span><br><span class="line">274</span><br><span class="line">275</span><br><span class="line">276</span><br><span class="line">277</span><br><span class="line">278</span><br><span class="line">279</span><br><span class="line">280</span><br><span class="line">281</span><br><span class="line">282</span><br><span class="line">283</span><br><span class="line">284</span><br><span class="line">285</span><br><span class="line">286</span><br><span class="line">287</span><br><span class="line">288</span><br><span class="line">289</span><br><span class="line">290</span><br><span class="line">291</span><br><span class="line">292</span><br><span class="line">293</span><br><span class="line">294</span><br><span class="line">295</span><br><span class="line">296</span><br><span class="line">297</span><br><span class="line">298</span><br><span class="line">299</span><br><span class="line">300</span><br><span class="line">301</span><br><span class="line">302</span><br><span class="line">303</span><br><span class="line">304</span><br><span class="line">305</span><br><span class="line">306</span><br><span class="line">307</span><br><span class="line">308</span><br><span class="line">309</span><br><span class="line">310</span><br><span class="line">311</span><br><span class="line">312</span><br><span class="line">313</span><br><span class="line">314</span><br><span class="line">315</span><br><span class="line">316</span><br><span class="line">317</span><br><span class="line">318</span><br><span class="line">319</span><br><span class="line">320</span><br><span class="line">321</span><br><span class="line">322</span><br><span class="line">323</span><br><span class="line">324</span><br><span class="line">325</span><br><span class="line">326</span><br><span class="line">327</span><br><span class="line">
230328</span><br><span class="line">329</span><br><span class="line">330</span><br><span class="line">331</span><br><span class="line">332</span><br><span class="line">333</span><br><span class="line">334</span><br><span class="line">335</span><br><span class="line">336</span><br><span class="line">337</span><br><span class="line">338</span><br><span class="line">339</span><br><span class="line">340</span><br><span class="line">341</span><br><span class="line">342</span><br><span class="line">343</span><br><span class="line">344</span><br><span class="line">345</span><br><span class="line">346</span><br><span class="line">347</span><br><span class="line">348</span><br><span class="line">349</span><br><span class="line">350</span><br><span class="line">351</span><br><span class="line">352</span><br><span class="line">353</span><br></pre></td><td class="code"><pre><span class="line">&quot;&quot;&quot;</span><br><span class="line">Some codes from https://github.com/Newmu/dcgan_code</span><br><span class="line">&quot;&quot;&quot;</span><br><span class="line">from __future__ import division</span><br><span class="line">from glob import glob</span><br><span class="line">from os.path import join,basename,exists</span><br><span class="line">from os import makedirs</span><br><span class="line">import math</span><br><span class="line">import json</span><br><span class="line">import random</span><br><span class="line">import pprint</span><br><span class="line">import scipy.misc</span><br><span class="line">import numpy as np</span><br><span class="line">from time import gmtime, strftime</span><br><span class="line">from six.moves import xrange</span><br><span class="line"></span><br><span class="line">import tensorflow as tf</span><br><span class="line">import tensorflow.contrib.slim as slim</span><br><span class="line"></span><br><span class="line">pp = pprint.PrettyPrinter()</span><br><span class="line"></span><br><span class="line">get_stddev = lambda x, k_h, k_w: 1/math.sqrt(k_w*k_h*x.get_shape()[-1])</span><br><span class="line"></span><br><span class="line">def show_all_variables():</span><br><span class="line">  model_vars = tf.trainable_variables()</span><br><span class="line">  # Prints the names and shapes of the variables</span><br><span class="line">  slim.model_analyzer.analyze_vars(model_vars, print_info=True)</span><br><span class="line"></span><br><span class="line">def get_image(image_path, input_height, input_width,</span><br><span class="line">              resize_height=64, resize_width=64,</span><br><span class="line">              crop=True, grayscale=False):</span><br><span class="line">  image = imread(image_path, grayscale)</span><br><span class="line">  return transform(image, input_height, input_width,</span><br><span class="line">                   resize_height, resize_width, crop)</span><br><span class="line"></span><br><span class="line">def save_images(images, size, image_path):</span><br><span class="line">  return imsave(inverse_transform(images), size, image_path)</span><br><span class="line"></span><br><span class="line">def imread(path, grayscale = False):</span><br><span class="line">  if (grayscale):</span><br><span class="line">    return scipy.misc.imread(path, flatten = True).astype(np.float)</span><br><span class="line">  else:</span><br><span class="line">    return scipy.misc.imread(path).astype(np.float)</span><br><span class="line"></span><br><span class="line">def merge_images(images, size):</span><br><span class="line">  return inverse_transform(images)</span><br><span class="line"></span><br><span class="line">def merge(images, size):</span><br><span class="line">  # samples 图片的真实高和宽</span><br><span class="line">  h, w = images.shape[1], images.shape[2]</span><br><span class="line">  # 图片channel的有效值只能是3或者4</span><br><span class="line">  if (images.shape[3] in (3,4)):</span><br><span class="line">    c = images.shape[3]</span><br><span class="line">    # img是合并之后的大图片,图片宽和高都倍增了</span><br><span class="line">    img = np.zeros((h * size[0], w * size[1], c))</span><br><span class="line">    # 遍历每一张图片</span><br><span class="line">    for idx, image in enumerate(images):</span><br><span class="line">      i = idx % size[1]</span><br><span class="line">      j = idx // size[1]</span><br><span class="line">      # 依次向大图填充小图(按行填充)</span><br><span class="line">      img[j * h:j * h + h, i * w:i * w + w, :] = image</span><br><span class="line">    return img</span><br><span class="line">  elif images.shape[3]==1:</span><br><span class="line">    # drop channel</span><br><span class="line">    img = np.zeros((h * size[0], w * size[1]))</span><br><span class="line">    for idx, image in enumerate(images):</span><br><span class="line">      i = idx % size[1]</span><br><span class="line">      j = idx // size[1]</span><br><span class="line">      img[j * h:j * h + h, i * w:i * w + w] = image[:,:,0]</span><br><span class="line">    return img</span><br><span class="line">  else:</span><br><span class="line">    raise ValueError(&apos;in merge(images,size) images parameter &apos;</span><br><span class="line">                     &apos;must have dimensions: HxW or HxWx3 or HxWx4&apos;)</span><br><span class="line"></span><br><span class="line">def imsave(images, size, path):</span><br><span class="line">  &apos;&apos;&apos;</span><br><span class="line">
230  modified imsave</span><br><span class="line">  :param images: ndarray,shape:(batch,height,width,channel)</span><br><span class="line">  :param size: (row images num,col images num)</span><br><span class="line">  :param path: save path</span><br><span class="line">  :return:</span><br><span class="line">  &apos;&apos;&apos;</span><br><span class="line">  # np.squeeze:去除维度为1的维</span><br><span class="line">  image = np.squeeze(merge(images, size))</span><br><span class="line">  return scipy.misc.imsave(path, image)</span><br><span class="line"></span><br><span class="line">def center_crop(x, crop_h, crop_w,</span><br><span class="line">                resize_h=64, resize_w=64):</span><br><span class="line">  &apos;&apos;&apos;</span><br><span class="line">  对图像进行中心化crop处理</span><br><span class="line">  :param x: image ndarray</span><br><span class="line">  :param crop_h: input height</span><br><span class="line">  :param crop_w: input width</span><br><span class="line">  :param resize_h: resize height</span><br><span class="line">  :param resize_w: resize width</span><br><span class="line">  :return: resized image</span><br><span class="line">  &apos;&apos;&apos;</span><br><span class="line">  if crop_w is None:</span><br><span class="line">    crop_w = crop_h</span><br><span class="line">  h, w = x.shape[:2]</span><br><span class="line">  j = int(round((h - crop_h)/2.))</span><br><span class="line">  i = int(round((w - crop_w)/2.))</span><br><span class="line">  return scipy.misc.imresize(</span><br><span class="line">      x[j:j+crop_h, i:i+crop_w], [resize_h, resize_w])</span><br><span class="line"></span><br><span class="line">def transform(image, input_height, input_width, </span><br><span class="line">              resize_height=64, resize_width=64, crop=True):</span><br><span class="line">  &apos;&apos;&apos;</span><br><span class="line">  对图像进行转化处理</span><br><span class="line">  :param image: ndarray of image</span><br><span class="line">  :param input_height: image height</span><br><span class="line">  :param input_width:  image width</span><br><span class="line">  :param resize_height: height after resize</span><br><span class="line">  :param resize_width:  width after resize</span><br><span class="line">  :param crop: if to crop or not</span><br><span class="line">  :return:</span><br><span class="line">  &apos;&apos;&apos;</span><br><span class="line">  if crop:</span><br><span class="line">    cropped_image = center_crop(</span><br><span class="line">      image, input_height, input_width, </span><br><span class="line">      resize_height, resize_width)</span><br><span class="line">  else:</span><br><span class="line">    # 直接resize</span><br><span class="line">    cropped_image = scipy.misc.imresize(image, [resize_height, resize_width])</span><br><span class="line">  # 将(0,255)映射到(-1,1)</span><br><span class="line">  return np.array(cropped_image)/127.5 - 1.</span><br><span class="line"></span><br><span class="line">def inverse_transform(images):</span><br><span class="line">  # (-1,1) ---&amp;gt; (0,1)</span><br><span class="line">  return (images+1.)/2.</span><br><span class="line"></span><br><span class="line">def to_json(output_path, *layers):</span><br><span class="line">  with open(output_path, &quot;w&quot;) as layer_f:</span><br><span class="line">    lines = &quot;&quot;</span><br><span class="line">    for w, b, bn in layers:</span><br><span class="line">      layer_idx = w.name.split(&apos;/&apos;)[0].split(&apos;h&apos;)[1]</span><br><span class="line"></span><br><span class="line">      B = b.eval()</span><br><span class="line"></span><br><span class="line">      if &quot;lin/&quot; in w.name:</span><br><span class="line">        W = w.eval()</span><br><span class="line">        depth = W.shape[1]</span><br><span class="line">      else:</span><br><span class="line">        W = np.rollaxis(w.eval(), 2, 0)</span><br><span class="line">        depth = W.shape[0]</span><br><span class="line"></span><br><span class="line">      biases = &#123;&quot;sy&quot;: 1, &quot;sx&quot;: 1, &quot;depth&quot;: depth, &quot;w&quot;: [&apos;%.2f&apos; % elem for elem in list(B)]&#125;</span><br><span class="line">      if bn != None:</span><br><span class="line">        gamma = bn.gamma.eval()</span><br><span class="line">        beta = bn.beta.eval()</span><br><span class="line"></span><br><span class="line">        gamma = &#123;&quot;sy&quot;: 1, &quot;sx&quot;: 1, &quot;depth&quot;: depth, &quot;w&quot;: [&apos;%.2f&apos; % elem for elem in list(gamma)]&#125;</span><br><span class="line">        beta = &#123;&quot;sy&quot;: 1, &quot;sx&quot;: 1, &quot;depth&quot;: depth, &quot;w&quot;: [&apos;%.2f&apos; % elem for elem in list(beta)]&#125;</span><br><span class="line">      else:</span><br><span class="line">        gamma = &#123;&quot;sy&quot;: 1, &quot;sx&quot;: 1, &quot;depth&quot;: 0, &quot;w&quot;: []&#125;</span><br><span class="line">        beta = &#123;&quot;sy&quot;: 1, &quot;sx&quot;: 1, &quot;depth&quot;: 0, &quot;w&quot;: []&#125;</span><br><span class="line"></span><br><span class="line">      if &quot;lin/&quot; in w.name:</span><br><span class="line">        fs = []</span><br><span class="line">        for w in W.T:</span><br><span class="line">          fs.append(&#123;&quot;sy&quot;: 1, &quot;sx&quot;: 1, &quot;depth&quot;: W.shape[0], &quot;w&quot;: [&apos;%.2f&apos; % elem for elem in list(w)]&#125;)</span><br><span class="line"></span><br><span class="line">        lines += &quot;&quot;&quot;</span><br><span class="line">          var layer_%s = &#123;</span><br><span class="line">            &quot;layer_type&quot;: &quot;fc&quot;, </span><br><span class="line">            &quot;sy&quot;: 1, &quot;sx&quot;: 1, </span><br><span class="line">            &quot;out_sx&quot;: 1, &quot;out_sy&quot;: 1,</span><br><span class="line">            &quot;
230stride&quot;: 1, &quot;pad&quot;: 0,</span><br><span class="line">            &quot;out_depth&quot;: %s, &quot;in_depth&quot;: %s,</span><br><span class="line">            &quot;biases&quot;: %s,</span><br><span class="line">            &quot;gamma&quot;: %s,</span><br><span class="line">            &quot;beta&quot;: %s,</span><br><span class="line">            &quot;filters&quot;: %s</span><br><span class="line">          &#125;;&quot;&quot;&quot; % (layer_idx.split(&apos;_&apos;)[0], W.shape[1], W.shape[0], biases, gamma, beta, fs)</span><br><span class="line">      else:</span><br><span class="line">        fs = []</span><br><span class="line">        for w_ in W:</span><br><span class="line">          fs.append(&#123;&quot;sy&quot;: 5, &quot;sx&quot;: 5, &quot;depth&quot;: W.shape[3], &quot;w&quot;: [&apos;%.2f&apos; % elem for elem in list(w_.flatten())]&#125;)</span><br><span class="line"></span><br><span class="line">        lines += &quot;&quot;&quot;</span><br><span class="line">          var layer_%s = &#123;</span><br><span class="line">            &quot;layer_type&quot;: &quot;deconv&quot;, </span><br><span class="line">            &quot;sy&quot;: 5, &quot;sx&quot;: 5,</span><br><span class="line">            &quot;out_sx&quot;: %s, &quot;out_sy&quot;: %s,</span><br><span class="line">            &quot;stride&quot;: 2, &quot;pad&quot;: 1,</span><br><span class="line">            &quot;out_depth&quot;: %s, &quot;in_depth&quot;: %s,</span><br><span class="line">            &quot;biases&quot;: %s,</span><br><span class="line">            &quot;gamma&quot;: %s,</span><br><span class="line">            &quot;beta&quot;: %s,</span><br><span class="line">            &quot;filters&quot;: %s</span><br><span class="line">          &#125;;&quot;&quot;&quot; % (layer_idx, 2**(int(layer_idx)+2), 2**(int(layer_idx)+2),</span><br><span class="line">               W.shape[0], W.shape[3], biases, gamma, beta, fs)</span><br><span class="line">    layer_f.write(&quot; &quot;.join(lines.replace(&quot;&apos;&quot;,&quot;&quot;).split()))</span><br><span class="line"></span><br><span class="line">def make_gif(images, fname, duration=2, true_image=False):</span><br><span class="line">  # 生成gif图</span><br><span class="line">  # duration:持续时间</span><br><span class="line">  # images shape:(batch_size,height,width,channel)</span><br><span class="line">  import moviepy.editor as mpy</span><br><span class="line"></span><br><span class="line">  def make_frame(t):</span><br><span class="line">    try:</span><br><span class="line">      # x 代表是t时刻选取的帧图片</span><br><span class="line">      x = images[int(len(images)/duration*t)]</span><br><span class="line">    except:</span><br><span class="line">      x = images[-1]</span><br><span class="line"></span><br><span class="line">    if true_image: # 返回不经过处理的ndarray,元素值是(-1,1)之间</span><br><span class="line">      return x.astype(np.uint8)</span><br><span class="line">    else:</span><br><span class="line">      # (-1,1) ---&amp;gt; (0,255)</span><br><span class="line">      return ((x+1)/2*255).astype(np.uint8)</span><br><span class="line"></span><br><span class="line">  clip = mpy.VideoClip(make_frame, duration=duration)</span><br><span class="line">  clip.write_gif(fname, fps = len(images) / duration)</span><br><span class="line"></span><br><span class="line">def visualize(sess, dcgan, config, option):</span><br><span class="line">  # 用于可视化</span><br><span class="line">  image_frame_dim = int(math.ceil(config.batch_size**.5)) # 图片尺寸</span><br><span class="line">  if option == 0:</span><br><span class="line">    # noise</span><br><span class="line">    z_sample = np.random.uniform(-0.5, 0.5, size=(config.batch_size, dcgan.z_dim))</span><br><span class="line">    samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;)</span><br><span class="line">    save_images(samples, [image_frame_dim, image_frame_dim], &apos;./%s/test_%s.png&apos; % (config.sample_dir,strftime(&quot;%Y-%m-%d-%H-%M-%S&quot;, gmtime())))</span><br><span class="line">  elif option == 1: # 将samples生成大图</span><br><span class="line">    values = np.arange(0, 1, 1./config.batch_size)</span><br><span class="line">    for idx in xrange(dcgan.z_dim):</span><br><span class="line">      print(&quot; [*] %d&quot; % idx)</span><br><span class="line">      z_sample = np.random.uniform(-1, 1, size=(config.batch_size , dcgan.z_dim))</span><br><span class="line">      for kdx, z in enumerate(z_sample):</span><br><span class="line">        z[idx] = values[kdx]</span><br><span class="line"></span><br><span class="line">      if config.dataset == &quot;mnist&quot;:</span><br><span class="line">        # y是batch_size个0-9之间的随机数</span><br><span class="line">        y = np.random.choice(10, config.batch_size)</span><br><span class="line">        save_random_digits(y,image_frame_dim,image_frame_dim,&apos;./%s/test_arange_%s.txt&apos; % (config.sample_dir,idx))</span><br><span class="line">        y_one_hot = np.zeros((config.batch_size, 10))</span><br><span class="line">        y_one_hot[np.arange(config.batch_size), y] = 1</span><br><span class="line"></span><br><span class="line">        samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample, dcgan.y: y_one_hot&#125;)</span><br><span class="line">      else:</span><br><span class="line">        samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;)</span><br><span class="line"></span><br><span class="line">      save_images(samples, [image_frame_dim, image_frame_dim], &ap
230os;./%s/test_arange_%s.png&apos; % (config.sample_dir,idx))</span><br><span class="line">  elif option == 2:</span><br><span class="line">    values = np.arange(0, 1, 1./config.batch_size)</span><br><span class="line">    # idx是随机的</span><br><span class="line">    for idx in [random.randint(0, dcgan.z_dim - 1) for _ in xrange(dcgan.z_dim)]:</span><br><span class="line">      print(&quot; [*] %d&quot; % idx)</span><br><span class="line">      # z_dim:test_images_num</span><br><span class="line">      z = np.random.uniform(-0.2, 0.2, size=(dcgan.z_dim))</span><br><span class="line">      # np.tile:按照指定的维度将array重复</span><br><span class="line">      # z_sample shape:(batch_size,z_dim)</span><br><span class="line">      z_sample = np.tile(z, (config.batch_size, 1))</span><br><span class="line">      #z_sample = np.zeros([config.batch_size, dcgan.z_dim])</span><br><span class="line">      for kdx, z in enumerate(z_sample):</span><br><span class="line">        z[idx] = values[kdx]</span><br><span class="line"></span><br><span class="line">      if config.dataset == &quot;mnist&quot;:</span><br><span class="line">        y = np.random.choice(10, config.batch_size)</span><br><span class="line">        #save_random_digits(y, image_frame_dim, image_frame_dim, &apos;./%s/test_%s.txt&apos; % % (config.sample_dir,strftime(&quot;%Y-%m-%d-%H-%M-%S&quot;, gmtime())))</span><br><span class="line">        y_one_hot = np.zeros((config.batch_size, 10))</span><br><span class="line">        y_one_hot[np.arange(config.batch_size), y] = 1</span><br><span class="line"></span><br><span class="line">        samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample, dcgan.y: y_one_hot&#125;)</span><br><span class="line">      else:</span><br><span class="line">        samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;)</span><br><span class="line"></span><br><span class="line">      try:</span><br><span class="line">        make_gif(samples, &apos;./%s/test_gif_%s.gif&apos; % (config.sample_dir,idx))</span><br><span class="line">      except:</span><br><span class="line">        save_images(samples, [image_frame_dim, image_frame_dim], &apos;./%s/test_%s.png&apos; % (config.sample_dir,strftime(&quot;%Y-%m-%d-%H-%M-%S&quot;, gmtime())))</span><br><span class="line">  elif option == 3: # 不能是mnist,直接生成gif</span><br><span class="line">    values = np.arange(0, 1, 1./config.batch_size)</span><br><span class="line">    for idx in xrange(dcgan.z_dim):</span><br><span class="line">      print(&quot; [*] %d&quot; % idx)</span><br><span class="line">      z_sample = np.zeros([config.batch_size, dcgan.z_dim])</span><br><span class="line">      for kdx, z in enumerate(z_sample):</span><br><span class="line">        z[idx] = values[kdx]</span><br><span class="line"></span><br><span class="line">      samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;)</span><br><span class="line">      make_gif(samples, &apos;./%s/test_gif_%s.gif&apos; % (config.sample_dir,idx))</span><br><span class="line">  elif option == 4:</span><br><span class="line">    image_set = []</span><br><span class="line">    values = np.arange(0, 1, 1./config.batch_size)</span><br><span class="line"></span><br><span class="line">    for idx in xrange(dcgan.z_dim):</span><br><span class="line">      print(&quot; [*] %d&quot; % idx)</span><br><span class="line">      z_sample = np.zeros([config.batch_size, dcgan.z_dim])</span><br><span class="line">      for kdx, z in enumerate(z_sample): z[idx] = values[kdx]</span><br><span class="line"></span><br><span class="line">      image_set.append(sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;))</span><br><span class="line">      #make_gif(image_set[-1], &apos;./%s/test_gif_%s.gif&apos; % (config.sample_dir,idx))</span><br><span class="line"></span><br><span class="line">    # 合成一张大图gif(64张大图)</span><br><span class="line">    new_image_set = [merge(np.array([images[idx] for images in image_set]), [10, 10]) \</span><br><span class="line">        for idx in range(63, -1, -1)] # 63-0</span><br><span class="line">    make_gif(new_image_set, &apos;./%s/test_gif_merged.gif&apos; % config.sample_dir, duration=8)</span><br><span class="line"></span><br><span class="line">def save_random_digits(arr,height,width,save_path):</span><br><span class="line">  &apos;&apos;&apos;</span><br><span class="line">  将arr中数字保存到文件,按行保存,共有height行,width列</span><br><span class="line">  :param arr: ndarray</span><br><span class="line">  :param height: 行数</span><br><span class="line">  :param width: 列数</span><br><span class="line">  :param save_path: 保存文件地址</span><br><span class="line">  :return:</span><br><span class="line">  &apos;&apos;&apos;</span><br><span class="line">  with open(save_path,&quot;w&quot;) as f:</span><br><span class="line">    for i in range(height):</span><br><span class="line">      for j in range(width):</span><br><span class="line">        if j != width-1:</span><br><span class="line">          f.write(&quot;%d,&quot; % arr[i*width+j])</span><br><span class="line">        else:</span><br><span class="line">          f.write(&quot;%d\n&quot; % arr[i*width+j])</span><br><span class="line">  f.close()</span><br><span class="line"></span><br><span class="line"></span><br><span class="line"></span><br><span class="line">def image_manifold_size(num_images):</span><br><span class="line">  manifold_h = int(np.floor(np.sqrt(num_images)))</span><br><span class="line">  manifold_w = int(np.ceil(np.sqrt(num_images)))</span><br><span class="line">  assert manifold_h * manifold_w == num_images</span><br><span class="line">  return manifold_h, manifold_w</span><br><span class="line"></span><br><span class="line">def resize_imgs(imgs_path,size,save_dir):</span><br><span class="line">  &apos;&apos;&apos;</span><br><span class="line">  将imgs_path文件夹的所有图片都resize到size大小,并重新保存到save_dir</span><br><span class="line">  :param imgs_path: 原始图像文件夹地址</span><br><span class="line">  :param size: resize之后的图像大小</span><br><span class="line">  :param save_dir: resize之后的图像保存地址</span><br><span class="line">  :return:</span><br><span class="line">  &apos;&apos;&apos;</span><br><span class="line">  if not exists(save_dir):</span><br><span class="line">    makedirs(save_dir)</span><br><span class="line">  imgs = glob(imgs_path+&quot;*.jpg&quot;)</span><br><span class="line">  for i,img in enumerate(imgs,1):</span><br><span class="line">    try:</span><br><span class="line">      img_arr = scipy.misc.imread(img)</span><br><span class="line">      new_img = scipy.misc.imresize(img_arr,size)</span><br><span class="line">      scipy.misc.imsave(join(save_dir,basename(img)),new_img)</span><br><span class="line">    except Exception as e:</span><br><span class="line">      print(e)</span><br><span class="line">    if i % 100 == 0:</span><br><span class="line">      print(&quot;Resize and save %d images!&quot; % i)</span><br><span class="line">  print(&quot;Resize and save all %d images!&quot; % len(imgs))</span><br><span class="line"></span><br><span class="line"></span><br><span class="line"># if __name__ == &apos;__main__&apos;:</span><br><span class="line">#     imgs_path = &quot;data/images/&quot;</span><br><span class="line">#     save_dir = &quot;data/lsun_new/&quot;</span><br><span class="line">#     size = (108,108)</span><br><span class="line">#     resize_imgs(imgs_path,size,save_dir)</span><br></pre></td></tr></table></figure>
231<p>utils.py定义了很多有用的全局工具函数,可以直接被其他的脚本调用。</p>
232<ul>
233<li>glob库用来list 某一个文件夹下的files;os库用来操作路径和文件夹等;pprint用于美观打印;gtime和strftime有用格式化日期;scipy.misc包含了很多和图像相关的有用的函数。</li>
234<li>24-27行的<strong>show_all_variables</strong>函数,调用了 <span class="lang:default decode:true      crayon-inline ">slim.model_analyzer.analyze_vars(vars,print_info)</span> 函数来打印model所有variables的信息。</li>
235<li>39-43行的<strong>imread</strong>函数封装了 <span class="lang:python decode:true      crayon-inline ">scipy.misc.imread</span> 函数,该函数参数 <span class="lang:default decode:true      crayon-inline ">flatten = True</span> 表示将color layer 展平成一个single gray-scale layer。</li>
236<li>48-73行的<strong>merge</strong>函数用于从一系列小图产生大图,images[0]表示小图的个数,h=images[1]表示小图的高,w = images[2]表示小图的宽,x_h = size[0]表示最终大图height应该扩展的倍数,x_w = size[1]表示最终大图width应该扩展的倍数。该函数最终生成一个高为h*x_h,宽为w*
236x_w的大图。表示大图的高度方向包含x_h个小图,宽度方向包含x_w个小图。</li>
237<li>75-85行定义了保存图像的<strong>imsave</strong>函数。注意 <span class="lang:default decode:true      crayon-inline ">np.squeeze</span> 可以去除数组中维度为1的那些维(降维),与之相反的操作是 <span class="lang:default decode:true      crayon-inline ">np.expand_dims(arr,axis)</span> 函数,可以给指定的axis维度增加一维。</li>
238<li>87-104行的<strong>center_crop</strong>函数的作用是中心化剪切处理,同时对图像进行了resize操作。</li>
239<li>106-126行的<strong>transform</strong>函数,也是对图像进行center_crop(可选)以及resize操作,只不过它最后将image array的每个元素的取值范围从(0,255)映射到(-1,1),(-1,1)是tanh函数的取值范围。</li>
240<li>132-193行的<strong>to_json</strong>函数将各个layers结构保存到json文件,我们不用这个函数,就不细说了。</li>
241<li>195-215行的<strong>make_gif</strong>函数可以将生成的序列图像转换为gif图像,这里使用moviepy库来完成这个工作,关于moviepy的介绍和使用,可以参考我之前的一篇<a href="http://www.movieb2b.com/2018/05/24/%E4%BB%8B%E7%BB%8D%E4%B8%80%E4%B8%AApython%E8%A7%86%E9%A2%91%E5%A4%84%E7%90%86%E5%BA%93moviepy/" target="_blank" rel="noopener">文章</a>。</li>
242<li>217-298行的<strong>visualize</strong>用于测试阶段生成图像样本,可以是单个jpg格式的图像,也可以是gif图像,还可以是小图拼接成的大图。visualize函数通过option变量的取值(可以取0,1,2,3,4五个值)来控制以五种不同的方式保存结果。</li>
243</ul>
244<ol>
245<li><em>option=0</em>:这种情况只适用于dataset 不等于mnist的情况,直接将samples merge成一个大图,然后保存即可,其中大图共有batch_size张小图,每行和每列各有ceil(sqrt(batch_size))个;</li>
246<li><em>option=1</em>:这种情况和option=0类似,只是它考虑到了dataset为mnist的情况,如果是mnist,则会随机生成batch_size个digit labels,然后从generator生成相应的数字,最后拼接成一个大图,这里我自己定义了一个<strong>save_random_digits</strong>函数用于将每次随机生成的数字保存到txt文件中去,这样后续可以验证生成的数字图像是否是我们希望生成的;</li>
247<li><em>option=2</em>:这种情况下,不会生成一张大图,而是生成含有batch_size帧的gif图,默认时间是2s,如果生成gif失败,则会生成和option=1一样的大图;</li>
248<li><em>option=3</em>:不能是mnist数据集,生成和option=2一样的gif。</li>
249<li><em>option=4</em>:合成一张大图的gif,一共有batch_size个大图,每个大图由<strong>z_dim</strong>(生成样本数目)个小图组成。</li>
250</ol>
251<ul>
252<li>300-316行的<strong>save_random_digits</strong>函数是我自定义的函数,用于将随机数字保存到txt文件;</li>
253<li>最后326-346行的<strong>resize_imgs</strong>函数是我自己添加的,作用就是将指定文件夹下的图像resize成指定的大小,这样我们就可以利用自己的数据集训练model了。</li>
254</ul>
255<h3 id="3-代码运行结果-生成图像效果验证"><a href="#3-代码运行结果-生成图像效果验证" class="headerlink" title="3. 代码运行结果(生成图像效果验证)"></a>3. 代码运行结果(生成图像效果验证)</h3><h4 id="1-mnist"><a href="#1-mnist" class="headerlink" title="1. mnist"></a>1. <strong>mnist</strong></h4><h4 id="nbsp-nbsp-nbsp-nbsp-根据我们上面的解读,运行如下命令即可以使用mnist训练DCGAN"><a href="#nbsp-nbsp-nbsp-nbsp-根据我们上面的解读,运行如下命令即可以使用mnist训练DCGAN" class="headerlink" title="&nbsp;&nbsp;&nbsp;&nbsp;根据我们上面的解读,运行如下命令即可以使用mnist训练DCGAN:"></a>&nbsp;&nbsp;&nbsp;&nbsp;根据我们上面的解读,运行如下命令即可以使用mnist训练DCGAN:</h4><pre class="lang:sh decode:true " title="train_mnist.sh">python3 main.py --dataset=mnist --input_height=28 --output_height=28 --train True
256</pre> 
257你需要确保main.py目录下的data/mnist文件夹下有已经解压缩的mnist数据文件。由于mnist数据规模不大,所以使用gpu训练大概只需要几十分钟。训练完成之后,训练过程中采样得到的生成图片保存在samples文件夹下,第一次采样和最后一次采样得到图片分别为下图1和图2所示:
258<div align="center">
259<img src="http://t1.aixinxi.net/o_1ced498bonb91jvabu5m82rufa.png-w.jpg">
260</div>
261<center>图1 mnist训练第一次采样生成图片</center>
262<div align="center">
263<img src="http://t1.aixinxi.net/o_1ced735uau75sq1i2l1cf11n30a.png-w.jpg">
264</div>
265<center>图2 mnist训练最后一次采样生成图片</center>
266可以看出随着训练的进行,生成的手写数字的质量确实是慢慢提高的。好了,接着利用训练得到的checkpoint来进行test,这里visualize的option参数设置为1,然后运行如下的命令即可以进行测试:
267
268<pre class="lang:sh decode:true ">python3 main.py --dataset=mnist --input_height=28 --output_height=28 --train False</pre> 
269测试默认会生成100张合成的大图,我们随机抽取一张,比如第66张吧,其真实的随机数字排列和生成的手写数字如下图3和图4所示:
270<div align="center">
271<img src="http://t1.aixinxi.net/o_1ced7emnkgmouab1fjk149t1amva.png-w.jpg">
272</div>
273<center>图3 第66张真实的随机数字排列</center>
274
275<div align="center">
276<img src="http://t1.aixinxi.net/o_1ced7npdg6s1127g1aau8eg1vr9a.png-w.jpg">
277</div>
278<center>图4 第66张生成的手写数字</center>
279可以发现生成的手写数字和真实的数字是完全符合的,通过随机查看其他的生成图片,可以发现基本全部是100%符合的,这说明conditional DCGAN是非常有效的。
280
281#### 2. **celebA**
282#### &nbsp;&nbsp;&nbsp;&nbsp;celebA数据集比mnist数据集规模要大,有大约20w+的人脸图片,图片是彩色的108*108尺寸。运行下面的命令即可以进行训练:
283
284<pre class="lang:sh decode:true " title="train_celebA">python3 main.py --dataset celebA --input_height=108 --crop --train True \
285                --epoch 2 --sample_dir ./celebA_samples --visualize True </pre> 
286注意默认训练采样保存的文件夹是samples文件夹,由于我们已经把mnist的结果保存在那里了,如果继续使用这个文件夹,celebA的结果会把之前的文件覆盖掉。为了避免这样的情况,我们重新设定保存sample的文件夹为celebA_samples文件夹,这个文件夹会在运行过程中自动创建,不需要手动创建。由于celebA的数据集规模较大,我电脑的配置是:ubuntu 16.04,tensorflow1.4.1,cuda8+cudnn6,显卡是nvidia GTX950M,显存4G。在batch_size = 64的情况下,大概1.5s可以训练一个batch,因此如果按照默认配置epoch=25,一个epoch的batch_num = ceil(202602/64)=3166,因此全部训练完大约需要的时间为1.5\*3166\*25/3600 ≈33h。由于我没有台式机,自己的笔记本不太可能一直训练这么长时间;机房的电脑配置太渣,train不动。所以我只能随便train一下了。我甚至一轮都没有训练完就停下来了。第1个epoch第100个batch生成的图像如下图5所示:
287<div align="center">
288<img src="http://t1.aixinxi.net/o_1ced8totluqs4nn15oqhuc3a.png-w.jpg">
289</div>
290<center>图5 第1个epoch第100个batch生成的图像</center>
291第1个epoch第2500个batch生成的图像如下图6所示:
292<div align="center">
293<img src="http://t1.aixinxi.net/o_1ced90si8stn1pontvt128u1tp6a.png-w.jpg">
294</div>
295<center>图6 第1个epoch第2500个batch生成的图像</center>
296可以发现,虽然都没有完整的训练一个epoch,但是第2500个batch生成的图像效果已经能初步看出人脸的轮廓了,如果你有足够的算力,不妨试着完整训练一下,最后得到的结果应该会相当不错。
297
298接着我们可以利用上面那个只训练了一点点的模型进行测试,测试celebA运行命令:
299
300<pre class="lang:sh decode:true ">python3 main.py --dataset celebA --input_height=108 --crop --train False \
301                --checkpoint_dir ./checkpoint --sample_dir ./celebA_samples</pre> 
302当然你仍然可以通过设定option的值来控制test的输出。下面的图7和图8是生成的gif图(图8由于体积太大已经转为jpg格式),由于训练非常不充分,因此效果不佳,但是仍然有脸部的轮廓:
303<div align="center">
304<img src="http://www.movieb2b.com/wp-content/uploads/2018/05/test_gif_66.gif">
305</div>
306<center>图7 celebA训练不到一轮生成脸部图像gif(小图)</center>
307<div align="center">
308<img src="http://t1.aixinxi.net/o_1cedacdggvrgdhhdac1atjdaja.jpg-w.jpg">
309</div>
310<center>图8 celebA训练不到一轮生成脸部图像gif(大图)</center>
311
312#### 3. **lsun**
313#### &nbsp;&nbsp;&nbsp;&nbsp;由于我使用download.py下载的lsun文件体积非常大(46G),而且格式是mdb格式的,不好直接读取。所以我后来从lsun的官网又自己重新下载了一个2G的图像压缩文件,解压缩之后大概有9000张图像,里面的图像种类较多,主要是关于各种自然景观的。由于图像数量不大,而且各个图像风格差异较大,因此不是很适合训练DCGAN(当然也是可以train的),所以我自己就没有实验了。如果大家有兴趣可以自己尝试训练一下看看效果怎么样。
314
315#### 4. **beauty_girls**
316#### &nbsp;&nbsp;&nbsp;&nbsp;这个是我自己搜集的数据集,看名字就知道是关于美女的啊。大约有2000张美女图,基本上是全身图,原图尺寸较大,而且size不统一,我们需要利用上面提到的utils.py中的resize_imgs函数首先将所有图片resize到相同的尺寸(这里我resize到width和height都是108),然后保存到文件夹beauty_girls,将该文件夹放入data目录下,然后运行如下的命令就可以训练:
317
318<pre class="lang:sh decode:true ">python3 main.py --dataset beauty_girls --input_height=108 --crop --train True \
319                --epoch 500 --sample_dir ./beauty_girls_samples --visualize True \
320                --print_every 10 --checkpoint_every 240
321</pre> 
322
323<p>这一次因为图片数量只有2000,所以我设定要训练500轮,我在晚上睡觉的时候用笔记本跑了一下,这下却翻车了,训练采样得到的图片是这样的:</p>
324<p><div align="center"><br><img src="http://t1.aixinxi.net/o_1cedbrmoblmmjm0kfe1hagfp3a.jpg-w.jpg"><br></div></p>
325<center>图9 beauty_girls 从上到下依次训练1è½®,66è½®,200è½®,300è½®,500轮生成的图像</center><br>可以发现从第1轮到第300轮生成图片的质量是提高的,但是再往后训练,特别是到了最后500轮的时候,图像明æ
325˜¾èŠ±äº†ï¼Œå¾ˆå¤šå°å›¾éƒ½æ˜¯ç›¸ä¼¼çš„çœ‹ä¸æ‡‚çš„æ¨¡å¼(也就是论文里说的mode collapse),这说明最多训练到300轮左右模型就已经差不多收敛了,再往后效果可能会更差,也许会发生mode collapse这种现象。这一点和论文最后提到的是一致的。而且可以发现即使是最好的生成图片,质量也不是特别好,这可能主要是与训练样本数太少(只有2000)而且图像风格差异太大引起的。最后,不要问我要原始训练图片,是拿什么图片训练的,你看生成图片难道猜不到么?哈哈哈。<br><br>#### 5. <strong>girl_face</strong><br>#### &nbsp;&nbsp;&nbsp;&nbsp;这个数据集来自知乎网友Best July的文章:<a href="https://zhuanlan.zhihu.com/p/26639135" target="_blank" rel="noopener">用DCGAN生成女朋友</a>,有兴趣大家可以看看这篇文章。该数据集包含了剪切好的8000多张妹子的头像,大小都是96x96的。差不多是下面这种:<br><div align="center"><br><img src="http://t1.aixinxi.net/o_1ceddckoak751cticps71vu4ca.jpg-w.jpg"><br></div><br><center>图10 girl_face 训练示例图片</center><br>数据集大家可以去<a href="https://pan.baidu.com/s/1dERYUmH" target="_blank" rel="noopener">faces</a>下载,密码:09h9。运行下面的命令即可以开始训练:<br><br><pre class="lang:sh decode:true ">python3 main.py –dataset girl_face –input_height=96 –crop –train True \<br>                –epoch 200 –sample_dir ./girl_face –visualize True \<br>                –print_every 30 –checkpoint_every 300</pre><br>你需要确保将包含图片数据的girl_face文件夹放在data目录下,我们设定训练200轮,全部训练完成估计要5,6个小时。下图11(从上至下)是分别训练1è½®,30è½®,70轮,100轮,130轮以及170轮时候产生的图像,可以发现随着训练轮数的增加,生成图像的质量是逐渐增加的,大概到100轮左右的时候,其实生成的头像质量已经很不错了(可以发现是美女了),后续个别位置的小图质量有所增加,但是始终有一些小图有一些畸变,不是特别自然。但是总体上来说,生成的图片质量很不错了。<br><div align="center"><br><img src="http://t1.aixinxi.net/o_1cedk932717201vnr5rn1jjovrca.jpg-w.jpg"><br></div><br><center>图11 girl_face 训练1è½®,30è½®,70轮,100轮,130轮以及170轮时候产生的图像(从上至下)</center>
326
327<p>训练完成之后,我们使用训练得到的model进行test,但是其实有一个问题我们之前没有提到,那就是如果训练轮数设定的过多,那么最新的一个checkpoint加载得到的model未必是最优的,最优的可能在中间的某一个epoch。但是原代码只能加载最新的一个checkpoint,所以我们将model.py中的 <strong>load</strong> 函数修改如下:</p>
328<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br></pre></td><td class="code"><pre><span class="line"><span class="comment"># load checkpoints file</span></span><br><span class="line">  <span class="function"><span class="keyword">def</span> <span class="title">load</span><span class="params">(self, checkpoint_dir,checkpoint_name = None)</span>:</span></span><br><span class="line">    <span class="keyword">
328import</span> re</span><br><span class="line">    print(<span class="string">" [*] Reading checkpoints..."</span>)</span><br><span class="line">    checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir)</span><br><span class="line">    <span class="comment">#A CheckpointState if the state was available, None</span></span><br><span class="line">    <span class="comment"># otherwise</span></span><br><span class="line">    ckpt = tf.train.get_checkpoint_state(checkpoint_dir)</span><br><span class="line">    <span class="keyword">if</span> ckpt <span class="keyword">and</span> ckpt.model_checkpoint_path:</span><br><span class="line">      <span class="comment"># basename:Returns the final component of a pathname</span></span><br><span class="line">      ckpt_name = os.path.basename(ckpt.model_checkpoint_path)</span><br><span class="line">      <span class="keyword">if</span> checkpoint_name <span class="keyword">is</span> <span class="keyword">None</span>:</span><br><span class="line">        <span class="comment"># 加载最新的checkpoint</span></span><br><span class="line">        self.saver.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))</span><br><span class="line">      <span class="keyword">else</span>:</span><br><span class="line">        <span class="comment"># 加载指定的而不是最新的checkpoint</span></span><br><span class="line">        self.saver.restore(self.sess, os.path.join(checkpoint_dir, checkpoint_name))</span><br><span class="line">      counter = int(next(re.finditer(<span class="string">"(\d+)(?!.*\d)"</span>,ckpt_name)).group(<span class="number">0</span>))</span><br><span class="line">      <span class="keyword">if</span> checkpoint_name <span class="keyword">is</span> <span class="keyword">None</span>:</span><br><span class="line">        print(<span class="string">" [*] Success to read &#123;&#125;"</span>.format(ckpt_name))</span><br><span class="line">      <span class="keyword">else</span>:</span><br><span class="line">        print(<span class="string">" [*] Success to read &#123;&#125;"</span>.format(checkpoint_name))</span><br><span class="line">      <span class="keyword">return</span> <span class="keyword">True</span>, counter</span><br><span class="line">    <span class="keyword">else</span>:</span><br><span class="line">      print(<span class="string">" [*] Failed to find a checkpoint"</span>)</span><br><span class="line">      <span class="keyword">return</span> <span class="keyword">False</span>, <span class="number">0</span></span><br></pre></td></tr></table></figure>
329<p>主要的修改就是增加了一个<strong>checkpoint_name</strong>参数,用于指定特定的而不是最新的checkpoint file。同时我们增加了一个checkpoint_name命令行参数: <span class="lang:default decode:true  crayon-inline ">flags.DEFINE_string(“checkpoint_name”,None,”the name of the loaded checkpoint file,default is the lastest checkpoint”)</span> 用来指定checkpoint_name参数,默认值是None。</p>
330<p>另外还有一个问题就是,在train的时候sample的样本,输入噪声z是服从(-1,1)的均匀分布,而原代码的visualize函数在option=1,2,3,4的时候,sample不是通过(-1,1)的均匀分布采样得到的,经过我的实验,如果在option=1,2,3,4的时候直接用原代码进行test,得到的生成图片几乎都是模糊的。我猜想这是因为test和train的时候的输入采样分布不一致导致的结果。因此我也对utils.py的visualize函数进行了修改如下:<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br><span class="line">44</span><br><span class="line">45</span><br><span class="line">46</span><br><span class="line">47</span><br><span class="line">48</span><br><span class="line">49</span><br><span class="line">50</span><br><span class="line">51</span><br><span class="line">52</span><br><span class="line">
33053</span><br><span class="line">54</span><br><span class="line">55</span><br><span class="line">56</span><br><span class="line">57</span><br><span class="line">58</span><br><span class="line">59</span><br><span class="line">60</span><br><span class="line">61</span><br><span class="line">62</span><br><span class="line">63</span><br><span class="line">64</span><br><span class="line">65</span><br><span class="line">66</span><br><span class="line">67</span><br><span class="line">68</span><br><span class="line">69</span><br><span class="line">70</span><br><span class="line">71</span><br><span class="line">72</span><br><span class="line">73</span><br><span class="line">74</span><br><span class="line">75</span><br><span class="line">76</span><br><span class="line">77</span><br><span class="line">78</span><br><span class="line">79</span><br><span class="line">80</span><br><span class="line">81</span><br><span class="line">82</span><br><span class="line">83</span><br><span class="line">84</span><br><span class="line">85</span><br><span class="line">86</span><br><span class="line">87</span><br><span class="line">88</span><br><span class="line">89</span><br><span class="line">90</span><br><span class="line">91</span><br><span class="line">92</span><br><span class="line">93</span><br><span class="line">94</span><br><span class="line">95</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">visualize</span><span class="params">(sess, dcgan, config, option)</span>:</span></span><br><span class="line">  <span class="comment"># 用于可视化</span></span><br><span class="line">  image_frame_dim = int(math.ceil(config.batch_size**<span class="number">.5</span>)) <span class="comment"># 图片尺寸</span></span><br><span class="line">  <span class="keyword">if</span> option == <span class="number">-1</span>:</span><br><span class="line">    <span class="comment"># noise</span></span><br><span class="line">    z_sample = np.random.uniform(<span class="number">-1</span>, <span class="number">1</span>, size=(config.batch_size, dcgan.z_dim))</span><br><span class="line">    samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;)</span><br><span class="line">    save_images(samples, [image_frame_dim, image_frame_dim],</span><br><span class="line">                <span class="string">'./%s/test_%s.png'</span> % (config.sample_dir, strftime(<span class="string">"%Y-%m-%d-%H-%M-%S"</span>, gmtime())))</span><br><span class="line">  <span class="keyword">elif</span> option == <span class="number">0</span>:</span><br><span class="line">    <span class="comment"># noise</span></span><br><span class="line">    z_sample = np.random.uniform(<span class="number">-0.5</span>, <span class="number">0.5</span>, size=(config.batch_size, dcgan.z_dim))</span><br><span class="line">    samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;)</span><br><span class="line">    save_images(samples, [image_frame_dim, image_frame_dim], <span class="string">'./%s/test_%s.png'</span> % (config.sample_dir,strftime(<span class="string">"%Y-%m-%d-%H-%M-%S"</span>, gmtime())))</span><br><span class="line">  <span class="keyword">elif</span> option == <span class="number">1</span>: <span class="comment"># 将samples生成大图</span></span><br><span class="line">    <span class="comment">#values = np.arange(0, 1, 1./config.batch_size)</span></span><br><span class="line">    <span class="keyword">for</span> idx <span class="keyword">in</span> xrange(dcgan.z_dim):</span><br><span class="line">      print(<span class="string">" [*] %d"</span> % idx)</span><br><span class="line">      z_sample = np.random.uniform(<span class="number">-1</span>, <span class="number">1</span>, size=(config.batch_size , dcgan.z_dim))</span><br><span class="line">      <span class="comment"># for kdx, z in enumerate(z_sample):</span></span><br><span class="line">      <span class="comment">#   z[idx] = values[kdx]</span></span><br><span class="line"></span><br><span class="line">
330      <span class="keyword">if</span> config.dataset == <span class="string">"mnist"</span>:</span><br><span class="line">        <span class="comment"># y是batch_size个0-9之间的随机数</span></span><br><span class="line">        y = np.random.choice(<span class="number">10</span>, config.batch_size)</span><br><span class="line">        save_random_digits(y,image_frame_dim,image_frame_dim,<span class="string">'./%s/test_arange_%s.txt'</span> % (config.sample_dir,idx))</span><br><span class="line">        y_one_hot = np.zeros((config.batch_size, <span class="number">10</span>))</span><br><span class="line">        y_one_hot[np.arange(config.batch_size), y] = <span class="number">1</span></span><br><span class="line"></span><br><span class="line">        samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample, dcgan.y: y_one_hot&#125;)</span><br><span class="line">      <span class="keyword">else</span>:</span><br><span class="line">        samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;)</span><br><span class="line"></span><br><span class="line">      save_images(samples, [image_frame_dim, image_frame_dim], <span class="string">'./%s/test_arange_%s.png'</span> % (config.sample_dir,idx))</span><br><span class="line">  <span class="keyword">elif</span> option == <span class="number">2</span>:</span><br><span class="line">    <span class="comment"># values = np.arange(0, 1, 1./config.batch_size)</span></span><br><span class="line">    <span class="comment"># idx是随机的</span></span><br><span class="line">    <span class="comment"># for idx in [random.randint(0, dcgan.z_dim - 1) for _ in xrange(dcgan.z_dim)]:</span></span><br><span class="line">    <span class="keyword">for</span> idx <span class="keyword">in</span> xrange(dcgan.z_dim):</span><br><span class="line">      print(<span class="string">" [*] %d"</span> % idx)</span><br><span class="line">      <span class="comment"># z_dim:test_images_num</span></span><br><span class="line">      <span class="comment">#z = np.random.uniform(-0.2, 0.2, size=(dcgan.z_dim))</span></span><br><span class="line">      <span class="comment"># np.tile:按照指定的维度将array重复</span></span><br><span class="line">      <span class="comment"># z_sample shape:(batch_size,z_dim)</span></span><br><span class="line">      <span class="comment">#z_sample = np.tile(z, (config.batch_size, 1))</span></span><br><span class="line">      <span class="comment">#z_sample = np.zeros([config.batch_size, dcgan.z_dim])</span></span><br><span class="line">      <span class="comment"># for kdx, z in enumerate(z_sample):</span></span><br><span class="line">      <span class="comment">#   z[idx] = values[kdx]</span></span><br><span class="line">      z_sample = np.random.uniform(<span class="number">-1</span>, <span class="number">1</span>, size=(config.batch_size, dcgan.z_dim))</span><br><span class="line">      <span class="keyword">if</span> config.dataset == <span class="string">"mnist"</span>:</span><br><span class="line">        y = np.random.choice(<span class="number">10</span>, config.batch_size)</span><br><span class="line">        <span class="comment">#save_random_digits(y, image_frame_dim, image_frame_dim, './%s/test_%s.txt' % % (config.sample_dir,strftime("%Y-%m-%d-%H-%M-%S", gmtime())))</span></span><br><span class="line">        y_one_hot = np.zeros((config.batch_size, <span class="number">10</span>))</span><br><span class="line">        y_one_hot[np.arange(config.batch_size), y] = <span class="number">1</span></span><br><span class="line"></span><br><span class="line">        samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample, dcgan.y: y_one_hot&#125;)</span><br><span class="line">      <span class="keyword">else</span>:</span><br><span class="line">        samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;)</span><br><span class="line">      <span class="keyword">try</span>:</span><br><span class="line">        make_gif(samples, <span class="string">'./%s/test_gif_%s.gif'</span> % (config.sample_dir,idx),<span class="number">4</span>)</span><br><span class="line">
330      <span class="keyword">except</span>:</span><br><span class="line">        save_images(samples, [image_frame_dim, image_frame_dim], <span class="string">'./%s/test_%s.png'</span> % (config.sample_dir,strftime(<span class="string">"%Y-%m-%d-%H-%M-%S"</span>, gmtime())))</span><br><span class="line">  <span class="keyword">elif</span> option == <span class="number">3</span>: <span class="comment"># 不能是mnist,直接生成gif</span></span><br><span class="line">    <span class="comment"># values = np.arange(0, 1, 1./config.batch_size)</span></span><br><span class="line">    <span class="keyword">for</span> idx <span class="keyword">in</span> xrange(dcgan.z_dim):</span><br><span class="line">      print(<span class="string">" [*] %d"</span> % idx)</span><br><span class="line">      <span class="comment"># z_sample = np.zeros([config.batch_size, dcgan.z_dim])</span></span><br><span class="line">      <span class="comment"># for kdx, z in enumerate(z_sample):</span></span><br><span class="line">      <span class="comment">#   z[idx] = values[kdx]</span></span><br><span class="line">      z_sample = np.random.uniform(<span class="number">-1</span>, <span class="number">1</span>, size=(config.batch_size, dcgan.z_dim))</span><br><span class="line">      samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;)</span><br><span class="line">      make_gif(samples, <span class="string">'./%s/test_gif_%s.gif'</span> % (config.sample_dir,idx),<span class="number">4</span>)</span><br><span class="line">  <span class="keyword">elif</span> option == <span class="number">4</span>:</span><br><span class="line">    image_set = []</span><br><span class="line">    <span class="comment"># values = np.arange(0, 1, 1./config.batch_size)</span></span><br><span class="line"></span><br><span class="line">    <span class="keyword">for</span> idx <span class="keyword">in</span> xrange(dcgan.z_dim):</span><br><span class="line">      print(<span class="string">" [*] %d"</span> % idx)</span><br><span class="line">      <span class="comment"># z_sample = np.zeros([config.batch_size, dcgan.z_dim])</span></span><br><span class="line">      <span class="comment"># for kdx, z in enumerate(z_sample): z[idx] = values[kdx]</span></span><br><span class="line">      z_sample = np.random.uniform(<span class="number">-1</span>, <span class="number">1</span>, size=(config.batch_size, dcgan.z_dim))</span><br><span class="line">      image_set.append(sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;))</span><br><span class="line">      make_gif(image_set[<span class="number">-1</span>], <span class="string">'./%s/test_gif_%s.gif'</span> % (config.sample_dir,idx),<span class="number">12</span>)</span><br><span class="line"></span><br><span class="line">    <span class="comment"># 合成一张大图gif(64张大图)</span></span><br><span class="line">    new_image_set = [merge(np.array([images[idx] <span class="keyword">for</span> images <span class="keyword">in</span> image_set]), [<span class="number">10</span>, <span class="number">10</span>]) \</span><br><span class="line">        <span class="keyword">for</span> idx <span class="keyword">in</span> range(<span class="number">63</span>, <span class="number">-1</span>, <span class="number">-1</span>)] <span class="comment"># 63-0</span></span><br><span class="line">    make_gif(new_image_set, <span class="string">'./%s/test_gif_merged.gif'</span> % config.sample_dir, duration=<span class="number">8</span>)</span><br><span class="line"></span><br><span class="line">  <span class="keyword">elif</span> option == <span class="number">5</span>:</span><br><span class="line">    <span class="comment">#保存单个的小图</span></span><br><span class="line">    z_sample = np.random.uniform(<span class="number">-1</span>, <span class="number">1</span>, size=(config.batch_size, dcgan.z_dim))</span><br><span class="line">    samples = sess.run(dcgan.sampler, feed_dict=&#123;dcgan.z: z_sample&#125;)</span><br><span class="line">    <span class="keyword">for</span> i,sample <span class="keyword">in</span> enumerate(samples):</span><br><span class="line">      scipy.misc.imsave(<span class="string">"./%s/single_test_%s.png"</span> %(config.sample_dir,i),sample)</span><br></pre></td></tr></table></figure></p>
331<p>主要的修改是将所有的采样方式都改为(-1,1)的均匀分布: <span class="lang:default decode:true    crayon-inline ">z_sample = np.random.uniform(-1, 1, size=(config.batch_size, dcgan.z_dim))</span> 。实验发现,这种方式在test的时候是非常有效的。另外,我保留了option=0的情况不变,增加了option=-1的情况以及option=5的情况。option=5表示将生成的图片按小图保存。下面的几张图展示了test的结果:</p>
332<p><div align="center"><br><img src="http://t1.aixinxi.net/o_1cedlm2ev1f516e11gb81vn5lgba.png-w.jpg"><br></div></p>
333<center>图12 girl_face 随机选取的一个test 生成图像大图</center>
334
335<p><div align="center"><br><img src="http://t1.aixinxi.net/o_1cedlt5tf12321k5k1rhnodqfroa.jpg-w.jpg"><br></div></p>
336<center>图13 girl_face 随机选取的几张test 生成图像小图合集</center>
337
338<p><div align="center"><br><img src="http://www.movieb2b.com/wp-content/uploads/2018/05/test_gif_66-1.gif" alt="" width="64" height="64" class="alignnone size-full wp-image-226"><br></div></p>
339<center>图14 girl_face 随机选取的生成gif图像</center>
340
341<p><div align="center"><br><img src="http://t1.aixinxi.net/o_1cedo45nicb8v9j1nsi10j6ttc.gif-j.jpg"><br></div></p>
342<center>图15 girl_face 生成的大图gif图像</center>
343
344<h3 id="4-总结"><a href="#4-总结" class="headerlink" title="4. 总结"></a>4. 总结</h3><h4 id="nbsp-nbsp-nbsp-nbsp-本文详细解读了DCGAN代码的tensorflow实现,并在mnist-celebA-以及自定义的数据集beauty-girs和girl-face数据集上进行了训练,测试。我们发现DCGAN确实在一定程度上提高了GAN训练的稳定性-不太容易发生mode-collapse的情况-而且生成的图片质量如果数据集数量较高、训练充分,还是很不错的。但是如果训练时间过长,还是可能会发生mode-collapse的情况,而且训练结果的质量也很取决于数据集的质量,数据集最好足够大-至少1w-吧-而且图片的风格最好是一致的,否则可能无法得到让人满意的结果-就像beauty-girls那样-。"><a href="#nbsp-nbsp-nbsp-nbsp-本文详细解读了DCGAN代码的tensorflow实现,并在mnist-celebA-以及自定义的数据集beauty-girs和girl-face数据集上进行了训练,测试。我们发现DCGAN确实在一定程度上提高了GAN训练的稳定性-不太容易发生mode-collapse的情况-而且生成的图片质量如果数据集数量较高、训练充分,还是很不错的。但是如果训练时间过长,还是可能会发生mode-collapse的情况,而且训练结果的质量也很取决于数据集的质量,数据集最好足够大-至少1w-吧-而且图片的风格最好是一致的,否则可能无法得到让人满意的结果-就像beauty-girls那样-。" class="headerlink" title="&nbsp;&nbsp;&nbsp;&nbsp;本文详细解读了DCGAN代码的tensorflow实现,并在mnist,celebA,以及自定义的数据集beauty_girs和girl_face数据集上进行了训练,测试。我们发现DCGAN确实在一定程度上提高了GAN训练的稳定性(不太容易发生mode collapse的情况),而且生成的图片质量如果数据集数量较高、训练充分,还是很不错的。但是如果训练时间过长,还是可能会发生mode collapse的情况,而且训练结果的质量也很取决于数据集的质量,数据集最好足够大(至少1w+吧),而且图片的风格最好是一致的,否则可能无法得到让人满意的结果(就像beauty_girls那样)。"></a>&nbsp;&nbsp;&nbsp;&nbsp;本文详细解读了DCGAN代码的tensorflow实现,并在mnist,celebA,以及自定义的数据集beauty_girs和girl_face数据集上进行了训练,测试。我们发现DCGAN确实在一定程度上提高了GAN训练的稳定性(不太容易发生mode collapse的情况),而且生成的图片质量如果数据集数量较高、训练充分,还是很不错的。但是如果训练时间过长,还是可能会发生mode collapse的情况,而且训练结果的质量也很取决于数据集的质量,数据集最好足够大(至少1w+吧),而且图片的风格最好是一致的,否则可能无法得到让人满意的结果(就像beauty_girls那样)。</h4><h2><center>本文完,感谢阅读!</center></h2>
345
346      
347
348      
349    </div>
350    <div class="article-info article-info-index">
351      
352      
353	<div class="article-tag tagcloud">
354		<i class="icon-price-tags icon"></i>
355		<ul class="article-tag-list">
356			 
357        		<li class="article-tag-list-item">
358        			<a href="javascript:void(0)" class="js-tag article-tag-list-link color4">GAN</a>
359        		</li>
360      		 
361        		<li class="article-tag-list-item">
362        			<a href="javascript:void(0)" class="js-tag article-tag-list-link color2">生成对抗网络</a>
363        		</li>
364      		 
365        		<li class="article-tag-list-item">
366        			<a href="javascript:void(0)" class="js-tag article-tag-list-link color1">DCGAN</a>
367        		</li>
368      		 
369        		<li class="article-tag-list-item">
370        			<a href="javascript:void(0)" class="js-tag article-tag-list-link color5">代码解读</a>
371        		</li>
372      		
373		</ul>
374	</div>
375
376      
377
378      
379        <p class="article-more-link">
380          <a class="article-more-a" href="/2018/05/27/DCGAN代码简单解读/">展开全文 >></a>
381        </p>
382      
383
384      
385      <div class="clearfix"></div>
386    </div>
387  </div>
388</article>
389
390<aside class="wrap-side-operation">
391    <div class="mod-side-operation">
392        
393        <div class="jump-container" id="js-jump-container" style="display:none;">
394            <a href="javascript:void(0)" class="mod-side-operation__jump-to-top">
395                <i class="icon-font icon-back"></i>
396            </a>
397            <div id="js-jump-plan-container" class="jump-plan-container" style="top: -11px;">
398                <i class="icon-font icon-plane jump-plane"></i>
399            </div>
400        </div>
401        
402        
403    </div>
404</aside>
405
406
407
408
409  
410    <article id="post-haskell简明入门-一" class="article article-type-post  article-index" itemscope itemprop="blogPost">
411  <div class="article-inner">
412    
413      <header class="article-header">
414        
415  
416    <h1 itemprop="name">
417      <a class="article-title" href="/2018/05/27/haskell简明入门-一/">haskell简明入门(一)</a>
418    </h1>
419  
420
421        
422        <a href="/2018/05/27/haskell简明入门-一/" class="archive-article-date">
423  	<time datetime="2018-05-27T09:08:37.000Z" itemprop="datePublished"><i class="icon-calendar icon"></i>2018-05-27</time>
424</a>
425        
426      </header>
427    
428    <div class="article-entry" itemprop="articleBody">
429      
430        <h3 id="本文的主要内容参考自《Haskell趣学指南》"><a href="#本文的主要内容参考自《Haskell趣学指南》" class="headerlink" title="本文的主要内容参考自《Haskell趣学指南》"></a>本文的主要内容参考自《Haskell趣学指南》</h3><p></p><h2>1. What is Haskell?</h2><br>&nbsp;&nbsp;&nbsp;&nbsp;以下内容引用自<a href="https://www.haskell.org/" target="_blank" rel="noopener">Haskell</a>官网:<p></p>
431<blockquote>
432<p>Haskell是一个先进的,纯粹的函数式编程语言。一个典型的声明式地,静态类型的代码如下:<br><figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br></pre></td><td class="code"><pre><span class="line"><span class="title">primes</span> = filterPrime [<span class="number">2.</span>.] </span><br><span class="line">  <span class="keyword">where</span> filterPrime (p:xs) = </span><br><span class="line">          p : filterPrime [x | x &lt;- xs, x `mod` p /= <span class="number">0</span>]</span><br></pre></td></tr></table></figure></p>
433</blockquote>
434<p>Haskell 有如下的特性:</p>
435<blockquote>
436<ul>
437<li><strong>静态类型</strong>(Statically Typed)。Haskell的每一个表达式都有一个在编译时决定的类型。所有的由函数引用组合起来的类型必须相匹配(match up),否则无法正常编译。类型不仅仅是形式上的保证,更是用于表达程序结构的语言。</li>
438<li><strong>纯粹函数式</strong>(Purely functional)。Haskell的每一个函数都是数学意义上的(pure)。即使是有副作用的IO操作也不过是在描述在做什么,同样由纯粹的代码产生。没有声明或者指令,仅仅只有表达式,该表达式不能对变量进行修改(局部变量或者全局变量),或者是获取像时间、随机数这样的状态。</li>
439<li><strong>类型推断</strong>(Type Inference)。你并不需要在Haskell中显式地写出每一种类型,类型将会进行双向推断。当然,你也可以选择自己写出类型,或者让编译为你进行推断。</li>
440<li><strong>并发</strong>(Concurrent)。Haskell可以很容易进行并发编程,这得力于它可以显式地处理effects。它的王牌编译器GHC带有一个高性能的并行的垃圾回收器,还有一个轻量级的并发库,其中包含了很多的有用的并发函数原型以及抽象接口。</li>
441<li><strong>惰性计算</strong>(Lazy)。函数并不会直接计算它们的值。这意味着程序可以在一起组合地非常好,可以通过只写通常的函数就表达控制结构(if/else)。Haskell代码的纯粹性使得它可以轻松将函数链式ç
441»„合,操作起来非常方便。</li>
442<li><strong>包</strong>(Packages)。你可以在public packages server 上找到非常多的活跃的开源Haskell packages。</li>
443</ul>
444</blockquote>
445<p></p><h2>2. How to use Haskell?</h2><br>&nbsp;&nbsp;&nbsp;&nbsp;你可以下载Haskell的GHC编译器,该编译器可以解释也可以编译Haskell程序。GHC还有一个很有用的交互模式,在终端输入ghci即可进入交互模式。然后运行命令<code>:l demo.hs</code>就可以加载demo.hs中的函数(Haskell程序文件以hs结尾)。如果你对demo.hs做了修改,可以重新运行命令<code>:l demo.hs</code>以重新加载其中的函数。这是我们以后实验学习Haskell的基本流程。<p></p>
446<p></p><h2> 3. Basic Operations</h2><p></p>
447<p></p><h3> 3.1 基本类型</h3><br>&nbsp;&nbsp;&nbsp;&nbsp;在终端输入ghci进行入交互模式。如下是一些基本的运算,几乎不用解释就可以看懂。<br><figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; <span class="number">2</span>+<span class="number">15</span></span><br><span class="line"><span class="number">17</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="number">45</span>*<span class="number">90</span></span><br><span class="line"><span class="number">4050</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="number">568</span><span class="number">-23</span></span><br><span class="line"><span class="number">545</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="number">3</span>/<span class="number">2</span></span><br><span class="line"><span class="number">1.5</span></span><br><span class="line"><span class="type">Prelude</span>&gt; (<span class="number">90</span><span class="number">-23</span>)*<span class="number">8</span></span><br><span class="line"><span class="number">536</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="number">50</span>*(<span class="number">100</span><span class="number">-9028</span>)</span><br><span class="line"><span class="number">-446400</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="type">True</span> &amp;&amp; <span class="type">False</span></span><br><span class="line"><span class="type">False</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="type">True</span> &amp;&amp; <span class="type">True</span></span><br><span class="line"><span class="type">True</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="type">False</span> || <span class="type">True</span></span><br><span class="line"><span class="type">True</span></span><br><span class="line"><span class="type">Prelude</span>&gt; not <span class="type">False</span></span><br><span class="line"><span class="type">True</span></span><br><span class="line"><span class="type">Prelude</span>&gt; not (<span class="type">True</span> &amp;&amp; <span class="type">True</span>)</span><br><span class="line"><span class="type">False</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="number">5</span> == <span class="number">4</span></span><br><span class="line"><span class="type">False</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="number">5</span> == <span class="number">5</span></span><br><span class="line"><span class="type">True</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="string">"Hello"</span> == <span class="string">"Hello"</span></span><br><span class="line"><span class="type">True</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="number">5</span> /= <span class="number">4</span></span><br><span class="line"><span class="type">True</span></span><br><span class="line"><span class="type">Prelude</span>&gt; <span class="number">5</span> /= <span class="number">5</span></span><br><span class="line"><span class="type">False</span></span><br></pre></td></tr></table></figure><p></p>
448<p></p><h3> 3.2 基础函数 </h3> <p></p>
449<h4 id="3-2-1-Haskell-常用内置函数"><a href="#3-2-1-Haskell-常用内置函数" class="headerlink" title="3.2.1 Haskell 常用内置函数"></a>3.2.1 Haskell 常用内置函数</h4><ul>
450<li><strong>succ</strong>。succ 函数返回一个数的后继。比如:<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; succ <span class="number">10</span></span><br><span class="line"><span class="number">11</span></span><br></pre></td></tr></table></figure>
451</li>
452</ul>
453<p>注意Haskell中函数的调用参数是使用空格的,这一点和传统的编译型语言有很大的不同。</p>
454<ul>
455<li><p><strong>
455pred</strong>。pred函数和succ函数相反,它返回一个数字的前驱。比如:</p>
456<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; pred <span class="number">10</span></span><br><span class="line"><span class="number">9</span></span><br></pre></td></tr></table></figure>
457</li>
458<li><p><strong>min</strong>。min函数接收两个数字参数,返回这两个数字的最小值。比如:</p>
459<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; min <span class="number">10</span> <span class="number">3.4</span></span><br><span class="line"><span class="number">3.4</span></span><br></pre></td></tr></table></figure>
460</li>
461</ul>
462<p>如果要比较多个数字的最小值,直接传递多个数字作为参数会报错,比如:<br><figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; min <span class="number">1</span> <span class="number">19</span> <span class="number">23</span></span><br><span class="line">&lt;interactive&gt;:<span class="number">21</span>:<span class="number">1</span>:</span><br><span class="line">    <span class="type">Non</span> <span class="class"><span class="keyword">type</span>-variable argument in the constraint: <span class="type">Num</span> (<span class="title">a</span> -&gt; <span class="title">t</span>)</span></span><br><span class="line">    (<span class="type">Use</span> <span class="type">FlexibleContexts</span> to permit this)</span><br><span class="line">    <span class="type">When</span> checking that ‘it’ has the inferred <span class="class"><span class="keyword">type</span></span></span><br><span class="line">      it :: <span class="keyword">forall</span> a t. (<span class="type">Num</span> a, <span class="type">Num</span> (a -&gt; t), <span class="type">Ord</span> (a -&gt; t)) =&gt; t</span><br></pre></td></tr></table></figure></p>
463<p>我们可以通过多次调用min函数(注意需要加括号)来解决:<br><figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; min <span class="number">1</span> (min <span class="number">9</span> <span class="number">23</span>)</span><br><span class="line"><span class="number">1</span></span><br></pre></td></tr></table></figure></p>
464<ul>
465<li><p><strong>max</strong>。max函数返回两个数字的最大值。比如:</p>
466<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; max <span class="number">9.9</span> <span class="number">9.0</span></span><br><span class="line"><span class="number">9.9</span></span><br></pre></td></tr></table></figure>
467</li>
468<li><p>Haskell中<strong>函数</strong>拥有最高的优先级。下面两句是等效的:</p>
469<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; succ <span class="number">10</span> + max <span class="number">10</span> <span class="number">20</span> * <span class="number">3</span></span><br><span class="line"><span class="number">71</span></span><br><span class="line"><span class="type">Prelude</span>&gt; (succ <span class="number">10</span>) +(max <span class="number">10</span> <span class="number">20</span>
469)*<span class="number">3</span></span><br><span class="line"><span class="number">71</span></span><br></pre></td></tr></table></figure>
470</li>
471</ul>
472<p>注意<code>succ 9*10</code>的值是100而不是91,这是因为计算的顺序是先计算<code>succ 9</code>得到10,然后再计算<code>10*10=100</code>,如果我们要计算9<em>10的后继,需要写成`succ (9</em>10)`的形式。</p>
473<ul>
474<li><strong>div</strong>。div函数计算一个整数除以另外一个整数的结果,比如:<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; div <span class="number">10</span> <span class="number">3</span></span><br><span class="line"><span class="number">3</span></span><br><span class="line"><span class="type">Prelude</span>&gt; div <span class="number">10</span> <span class="number">5</span></span><br><span class="line"><span class="number">2</span></span><br></pre></td></tr></table></figure>
475</li>
476</ul>
477<p>需要注意的是,div函数接收的两个参数必须是整数,第一个是被除数,第二个是除数,如果除不尽则返回商,舍弃余数。传入数字不能是浮点数,比如下面的形式会报错:<br><figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; div <span class="number">1.2</span> <span class="number">3</span></span><br><span class="line">&lt;interactive&gt;:<span class="number">31</span>:<span class="number">1</span>:</span><br><span class="line">    <span class="type">No</span> <span class="keyword">instance</span> for (<span class="type">Fractional</span> a0) arising from a use <span class="keyword">of</span> ‘it’</span><br><span class="line">    <span class="type">The</span> <span class="class"><span class="keyword">type</span> variable ‘a0’ is ambiguous</span></span><br><span class="line">    <span class="type">Note</span>: there are several potential instances:</span><br><span class="line"><span class="class">      <span class="keyword">instance</span> <span class="type">Integral</span> a =&gt; <span class="type">Fractional</span> (<span class="type">GHC</span>.<span class="type">Real</span>.<span class="type">Ratio</span> <span class="title">a</span>)</span></span><br><span class="line"><span class="class">        <span class="comment">-- Defined in ‘GHC.Real’</span></span></span><br><span class="line"><span class="class">      <span class="keyword">instance</span> <span class="type">Fractional</span> <span class="type">Double</span> <span class="comment">-- Defined in ‘GHC.Float’</span></span></span><br><span class="line"><span class="class">      <span class="keyword">instance</span> <span class="type">Fractional</span> <span class="type">Float</span> <span class="comment">-- Defined in ‘GHC.Float’</span></span></span><br><span class="line"><span class="class">    <span class="type">In</span> the first argument of ‘print’, namely ‘it’</span></span><br><span class="line"><span class="class">    <span class="type">In</span> a stmt of an interactive <span class="type">GHCi</span> command: print it</span></span><br></pre></td></tr></table></figure></p>
478<h4 id="3-2-2-构建自己的函数"><a href="#3-2-2-构建自己的函数" class="headerlink" title="3.2.2 构建自己的函数"></a>3.2.2 构建自己的函数</h4><ul>
479<li>doubleMe。doubleMe函数的定义如下:<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line"><span class="title">doubleMe</span> x = x + x <span class="comment">-- 将数字变为两倍</span></span><br></pre></td></tr></table></figure>
480</li>
481</ul>
482<p>这个函数非常简单,就是返回一个数字的两倍。注意上面函数的定义方式:首先是函数名字,然后是空格分隔的参数,接着是等于å
482·ï¼Œç­‰äºŽå·åŽé¢çš„æ˜¯å‡½æ•°çš„实现。要运行这个函数,需要把它写在一个文件里,比如add.hs,然后输入ghci进入命令行模式,输入<code>:l add.hs</code>就可以加载函数了。下面是它的调用计算结果:<br><figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br></pre></td><td class="code"><pre><span class="line"><span class="type">Prelude</span>&gt; :l add.hs </span><br><span class="line">[<span class="number">1</span> <span class="keyword">of</span> <span class="number">1</span>] <span class="type">Compiling</span> <span class="type">Main</span>             ( add.hs, interpreted )</span><br><span class="line"><span class="type">Ok</span>, modules loaded: <span class="type">Main</span>.</span><br><span class="line">*<span class="type">Main</span>&gt; doubleMe <span class="number">2</span></span><br><span class="line"><span class="number">4</span></span><br><span class="line">*<span class="type">Main</span>&gt; doubleMe <span class="number">3</span></span><br><span class="line"><span class="number">6</span></span><br></pre></td></tr></table></figure></p>
483<ul>
484<li><strong>doubleUs</strong>。该函数接收x,y两个参数,然后返回这两个参数之和的两倍。定义如下:<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line"><span class="title">doubleUs</span> x y = x*<span class="number">2</span> + y*<span class="number">2</span> <span class="comment">-- 两个数字变为两倍然后相加</span></span><br></pre></td></tr></table></figure>
485</li>
486</ul>
487<p>调用计算结果为:<br><figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br></pre></td><td class="code"><pre><span class="line">*<span class="type">Main</span>&gt; doubleUs <span class="number">1</span> <span class="number">3</span></span><br><span class="line"><span class="number">8</span></span><br><span class="line">*<span class="type">Main</span>&gt; doubleUs <span class="number">0</span> <span class="number">2</span></span><br><span class="line"><span class="number">4</span></span><br></pre></td></tr></table></figure></p>
488<ul>
489<li><strong>useDoubleMe</strong>。该函数功能和<code>doubleUs</code>一样,只不过其调用了<code>doubleMe</code>两次,代码如下:<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line"><span class="title">useDoubleMe</span> x y = doubleMe x + doubleMe y <span class="comment">-- 调用简单函数</span></span><br></pre></td></tr></table></figure>
490</li>
491</ul>
492<p>对了,忘了说了,haskell中注释是<code>--</code>。</p>
493<ul>
494<li><strong>doubleSmallNumber</strong>。该函数使用<code>if else</code>语句实现的功能是:如果x小于100则返回x;否则返回x的两倍。代码如下:<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line"><span class="title">doubleSmallNumber</span> x = <span class="keyword">if</span> x &lt; <span class="number">100</span> <span class="keyword">then</span> x <span class="keyword">else</span> <span class="number">2</span>*x <span class="comment">--如果x小于100返回x,否则返回2*x</span></span><br></pre></td></tr></table></figure>
495</li>
496</ul>
497<p>调用结果为:<br><figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br></pre></td><td class="code"><pre><span class="line">*<span class="type">Main</span>&gt; doubleSmallNumber <span class="number">20</span></span><br><span class="line"><span class="number">20</span></span><br><span class="line">*<span class="type">Main</span>&gt; doubleSmallNumber <span class="number">200</span></span><br><span class="line"><span class="number">400</span></span><br></pre></td></tr></table></figure></p>
498<p>需要注意的是Haskell中if和else一定是一起出现的,else不可以省略,而且本质上if和else都是需要返回一个值。Haskell中所有的函数和表达式都需要返回一个结果,if语句就是一个表达式,所以它一定会返回一个结果。</p>
499<ul>
500<li><strong>doubleSmallNumber’</strong>。doubleSmallNumber’函数(注意函数最后有一个<code>&#39;</code>)返回的结果是doubleSmallNumber返回值加1,定义为:<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line"><span class="title">doubleSmallNumber'</span> x = (<span class="keyword">if</span> x &lt; <span class="number">100</span> <span class="keyword">then</span> x <span class="keyword">else</span> <span class="number">2</span>*x) + <span class="number">1</span> <span class="comment">-- doubleSmallNumber返回值加1</span></span><br></pre></td></tr></table></figure>
501</li>
502</ul>
503<p>我们通常在Haskell中在函数名字最后加一个<code>&#39;</code>来表示对某个函数稍微修改之后得到的新函数。上面的函数如果把括号去掉,那么只会在<code>x&gt;=100</code>
503的时候加1了。另外需要注意的是,Haskell的函数名字开头不能是大写字母。</p>
504<ul>
505<li><strong>conanO’Brien</strong>。conanO’Brien是一个没有参数的函数,这样的函数也被称为”定义”或者”名字”。conanO’Brien定义如下:<figure class="highlight haskell"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line"><span class="title">conanO'Brien</span> = <span class="string">"It's a-me, Conan O'Brien!"</span> <span class="comment">-- 没有参数的函数</span></span><br></pre></td></tr></table></figure>
506</li>
507</ul>
508<p>上面的conanO’Brien函数定义好了之后,<code>conanO&#39;Brien</code>就与字符串<code>&quot;It&#39;s a-me, Conan O&#39;Brien!&quot;</code>等价了,而且字符串的值是不可以修改的。</p>
509
510      
511
512      
513    </div>
514    <div class="article-info article-info-index">
515      
516      
517	<div class="article-tag tagcloud">
518		<i class="icon-price-tags icon"></i>
519		<ul class="article-tag-list">
520			 
521        		<li class="article-tag-list-item">
522        			<a href="javascript:void(0)" class="js-tag article-tag-list-link color3">haskell</a>
523        		</li>
524      		
525		</ul>
526	</div>
527
528      
529
530      
531        <p class="article-more-link">
532          <a class="article-more-a" href="/2018/05/27/haskell简明入门-一/">展开全文 >></a>
533        </p>
534      
535
536      
537      <div class="clearfix"></div>
538    </div>
539  </div>
540</article>
541
542<aside class="wrap-side-operation">
543    <div class="mod-side-operation">
544        
545        <div class="jump-container" id="js-jump-container" style="display:none;">
546            <a href="javascript:void(0)" class="mod-side-operation__jump-to-top">
547                <i class="icon-font icon-back"></i>
548            </a>
549            <div id="js-jump-plan-container" class="jump-plan-container" style="top: -11px;">
550                <i class="icon-font icon-plane jump-plane"></i>
551            </div>
552        </div>
553        
554        
555    </div>
556</aside>
557
558
559
560
561  
562  
563    <nav id="page-nav">
564      <a class="extend prev" rel="prev" href="/">&laquo; Prev</a><a class="page-number" href="/">1</a><span class="page-number current">2</span>
565    </nav>
566  
567
568
569          </div>
570        </div>
571      </div>
572      <footer id="footer">
573  <div class="outer">
574    <div id="footer-info">
575    	<div class="footer-left">
576    		&copy; 2018 lyrichu
577    	</div>
578      	<div class="footer-right">
579      		<a href="http://hexo.io/" target="_blank">Hexo</a>  Theme <a href="https://github.com/litten/hexo-theme-yilia" target="_blank">Yilia</a> by Litten
580      	</div>
581    </div>
582  </div>
583</footer>
584    </div>
585    
585<script>
586	var yiliaConfig = {
587		mathjax: true,
588		isHome: true,
589		isPost: false,
590		isArchive: false,
591		isTag: false,
592		isCategory: false,
593		open_in_new: true,
594		toc_hide_index: true,
595		root: "/",
596		innerArchive: true,
597		showTags: true
598	}
599</script>
599
600
601<script>!function(t){function n(e){if(r[e])return r[e].exports;var i=r[e]={exports:{},id:e,loaded:!1};return t[e].call(i.exports,i,i.exports,n),i.loaded=!0,i.exports}var r={};n.m=t,n.c=r,n.p="./",n(0)}([function(t,n,r){r(195),t.exports=r(191)},function(t,n,r){var e=r(3),i=r(52),o=r(27),u=r(28),c=r(53),f="prototype",a=function(t,n,r){var s,l,h,v,p=t&a.F,d=t&a.G,y=t&a.S,g=t&a.P,b=t&a.B,m=d?e:y?e[n]||(e[n]={}):(e[n]||{})[f],x=d?i:i[n]||(i[n]={}),w=x[f]||(x[f]={});d&&(r=n);for(s in r)l=!p&&m&&void 0!==m[s],h=(l?m:r)[s],v=b&&l?c(h,e):g&&"function"==typeof h?c(Function.call,h):h,m&&u(m,s,h,t&a.U),x[s]!=h&&o(x,s,v),g&&w[s]!=h&&(w[s]=h)};e.core=i,a.F=1,a.G=2,a.S=4,a.P=8,a.B=16,a.W=32,a.U=64,a.R=128,t.exports=a},function(t,n,r){var e=r(6);t.exports=function(t){if(!e(t))throw TypeError(t+" is not an object!");return t}},function(t,n){var r=t.exports="undefined"!=typeof window&&window.Math==Math?window:"undefined"!=typeof self&&self.Math==Math?self:Function("return this")();"number"==typeof __g&&(__g=r)},function(t,n){t.exports=function(t){try{return!!t()}catch(t){return!0}}},function(t,n){var r=t.exports="undefined"!=typeof window&&window.Math==Math?window:"undefined"!=typeof self&&self.Math==Math?self:Function("return this")();"number"==typeof __g&&(__g=r)},function(t,n){t.exports=function(t){return"object"==typeof t?null!==t:"function"==typeof t}},function(t,n,r){var e=r(126)("wks"),i=r(76),o=r(3).Symbol,u="function"==typeof o;(t.exports=function(t){return e[t]||(e[t]=u&&o[t]||(u?o:i)("Symbol."+t))}).store=e},function(t,n){var r={}.hasOwnProperty;t.exports=function(t,n){return r.call(t,n)}},function(t,n,r){var e=r(94),i=r(33);t.exports=function(t){return e(i(t))}},function(t,n,r){t.exports=!r(4)(function(){return 7!=Object.defineProperty({},"a",{get:function(){return 7}}).a})},function(t,n,r){var e=r(2),i=r(167),o=r(50),u=Object.defineProperty;n.f=r(10)?Object.defineProperty:function(t,n,r){if(e(t),n=o(n,!0),e(r),i)try{return u(t,n,r)}catch(t){}if("get"in r||"set"in r)throw TypeError("Accessors not supported!");return"value"in r&&(t[n]=r.value),t}},function(t,n,r){t.exports=!r(18)(function(){return 7!=Object.defineProperty({},"a",{get:function(){return 7}}).a})},function(t,n,r){var e=r(14),i=r(22);t.exports=r(12)?function(t,n,r){return e.f(t,n,i(1,r))}:function(t,n,r){return t[n]=r,t}},function(t,n,r){var e=r(20),i=r(58),o=r(42),u=Object.defineProperty;n.f=r(12)?Object.defineProperty:function(t,n,r){if(e(t),n=o(n,!0),e(r),i)try{return u(t,n,r)}catch(t){}if("get"in r||"set"in r)throw TypeError("Accessors not supported!");return"value"in r&&(t[n]=r.value),t}},function(t,n,r){var e=r(40)("wks"),i=r(23),o=r(5).Symbol,u="function"==typeof o;(t.exports=function(t){return e[t]||(e[t]=u&&o[t]||(u?o:i)("Symbol."+t))}).store=e},function(t,n,r){var e=r(67),i=Math.min;t.exports=function(t){return t>0?i(e(t),9007199254740991):0}},function(t,n,r){var e=r(46);t.exports=function(t){return Object(e(t))}},function(t,n){t.exports=function(t){try{return!!t()}catch(t){return!0}}},function(t,n,r){var e=r(63),i=r(34);t.exports=Object.keys||function(t){return e(t,i)}},function(t,n,r){var e=r(21);t.exports=function(t){if(!e(t))throw TypeError(t+" is not an object!");return t}},function(t,n){t.exports=function(t){return"object"==typeof t?null!==t:"function"==typeof t}},function(t,n){t.exports=function(t,n){return{enumerable:!(1&t),configurable:!(2&t),writable:!(4&t),value:n}}},function(t,n){var r=0,e=Math.random();t.exports=function(t){return"Symbol(".concat(void 0===t?"":t,")_",(++r+e).toString(36))}},function(t,n){var r={}.hasOwnProperty;t.exports=function(t,n){return r.call(t,n)}},function(t,n){var r=t.exports={version:"2.4.0"};"number"==typeof __e&&(__e=r)},function(t,n){t.exports=function(t){if("function"!=typeof t)throw TypeError(t+" is not a function!");return t}},function(t,n,r){var e=r(11),i=r(66);t.exports=r(10)?function(t,n,r){return e.f(t,n,i(1,r))}:function(t,n,r){return t[n]=r,t}},function(t,n,r){var e=r(3),i=r(27),o=r(24),u=r(76)("src"),c="toString",f=Function[c],a=(""+f).split(c);r(52).inspectSource=function(t){return f.call(t)},(t.exports=function(t,n,r,c){var f="function"==typeof r;f&&(o(r,"name")||i(r,"name",n)),t[n]!==r&&(f&&(o(r,u)||i(r,u,t[n]?""+t[n]:a.join(String(n)))),t===e?t[n]=r:c?t[n]?t[n]=r:i(t,n,r):(delete t[n],i(t,n,r)))})(Function.prototype,c,function(){return"function"==typeof this&&this[u]||f.call(this)})}
601,function(t,n,r){var e=r(1),i=r(4),o=r(46),u=function(t,n,r,e){var i=String(o(t)),u="<"+n;return""!==r&&(u+=" "+r+'="'+String(e).replace(/"/g,"&quot;")+'"'),u+">"+i+"</"+n+">"};t.exports=function(t,n){var r={};r[t]=n(u),e(e.P+e.F*i(function(){var n=""[t]('"');return n!==n.toLowerCase()||n.split('"').length>3}),"String",r)}},function(t,n,r){var e=r(115),i=r(46);t.exports=function(t){return e(i(t))}},function(t,n,r){var e=r(116),i=r(66),o=r(30),u=r(50),c=r(24),f=r(167),a=Object.getOwnPropertyDescriptor;n.f=r(10)?a:function(t,n){if(t=o(t),n=u(n,!0),f)try{return a(t,n)}catch(t){}if(c(t,n))return i(!e.f.call(t,n),t[n])}},function(t,n,r){var e=r(24),i=r(17),o=r(145)("IE_PROTO"),u=Object.prototype;t.exports=Object.getPrototypeOf||function(t){return t=i(t),e(t,o)?t[o]:"function"==typeof t.constructor&&t instanceof t.constructor?t.constructor.prototype:t instanceof Object?u:null}},function(t,n){t.exports=function(t){if(void 0==t)throw TypeError("Can't call method on  "+t);return t}},function(t,n){t.exports="constructor,hasOwnProperty,isPrototypeOf,propertyIsEnumerable,toLocaleString,toString,valueOf".split(",")},function(t,n){t.exports={}},function(t,n){t.exports=!0},function(t,n){n.f={}.propertyIsEnumerable},function(t,n,r){var e=r(14).f,i=r(8),o=r(15)("toStringTag");t.exports=function(t,n,r){t&&!i(t=r?t:t.prototype,o)&&e(t,o,{configurable:!0,value:n})}},function(t,n,r){var e=r(40)("keys"),i=r(23);t.exports=function(t){return e[t]||(e[t]=i(t))}},function(t,n,r){var e=r(5),i="__core-js_shared__",o=e[i]||(e[i]={});t.exports=function(t){return o[t]||(o[t]={})}},function(t,n){var r=Math.ceil,e=Math.floor;t.exports=function(t){return isNaN(t=+t)?0:(t>0?e:r)(t)}},function(t,n,r){var e=r(21);t.exports=function(t,n){if(!e(t))return t;var r,i;if(n&&"function"==typeof(r=t.toString)&&!e(i=r.call(t)))return i;if("function"==typeof(r=t.valueOf)&&!e(i=r.call(t)))return i;if(!n&&"function"==typeof(r=t.toString)&&!e(i=r.call(t)))return i;throw TypeError("Can't convert object to primitive value")}},function(t,n,r){var e=r(5),i=r(25),o=r(36),u=r(44),c=r(14).f;t.exports=function(t){var n=i.Symbol||(i.Symbol=o?{}:e.Symbol||{});"_"==t.charAt(0)||t in n||c(n,t,{value:u.f(t)})}},function(t,n,r){n.f=r(15)},function(t,n){var r={}.toString;t.exports=function(t){return r.call(t).slice(8,-1)}},function(t,n){t.exports=function(t){if(void 0==t)throw TypeError("Can't call method on  "+t);return t}},function(t,n,r){var e=r(4);t.exports=function(t,n){return!!t&&e(function(){n?t.call(null,function(){},1):t.call(null)})}},function(t,n,r){var e=r(53),i=r(115),o=r(17),u=r(16),c=r(203);t.exports=function(t,n){var r=1==t,f=2==t,a=3==t,s=4==t,l=6==t,h=5==t||l,v=n||c;return function(n,c,p){for(var d,y,g=o(n),b=i(g),m=e(c,p,3),x=u(b.length),w=0,S=r?v(n,x):f?v(n,0):void 0;x>w;w++)if((h||w in b)&&(d=b[w],y=m(d,w,g),t))if(r)S[w]=y;else if(y)switch(t){case 3:return!0;case 5:return d;case 6:return w;case 2:S.push(d)}else if(s)return!1;return l?-1:a||s?s:S}}},function(t,n,r){var e=r(1),i=r(52),o=r(4);t.exports=function(t,n){var r=(i.Object||{})[t]||Object[t],u={};u[t]=n(r),e(e.S+e.F*o(function(){r(1)}),"Object",u)}},function(t,n,r){var e=r(6);t.exports=function(t,n){if(!e(t))return t;var r,i;if(n&&"function"==typeof(r=t.toString)&&!e(i=r.call(t)))return i;if("function"==typeof(r=t.valueOf)&&!e(i=r.call(t)))return i;if(!n&&"function"==typeof(r=t.toString)&&!e(i=r.call(t)))return i;throw TypeError("Can't convert object to primitive value")}},function(t,n,r){var e=r(5),i=r(25),o=r(91),u=r(13),c="prototype",f=function(t,n,r){var a,s,l,h=t&f.F,v=t&f.G,p=t&f.S,d=t&f.P,y=t&f.B,g=t&f.W,b=v?i:i[n]||(i[n]={}),m=b[c],x=v?e:p?e[n]:(e[n]||{})[c];v&&(r=n);for(a in r)(s=!h&&x&&void 0!==x[a])&&a in b||(l=s?x[a]:r[a],b[a]=v&&"function"!=typeof x[a]?r[a]:y&&s?o(l,e):g&&x[a]==l?function(t){var n=function(n,r,e){if(this instanceof t){switch(arguments.length){case 0:return new t;case 1:return new t(n);case 2:return new t(n,r)}return new t(n,r,e)}return t.apply(this,arguments)};return n[c]=t[c],n}(l):d&&"function"==typeof l?o(Function.call,l):l,d&&((b.virtual||(b.virtual={}))[a]=l,t&f.R&&m&&!m[a]&&u(m,a,l)))};f.F=1,f.G=2,f.S=4,f.P=8,f.B=16,f.W=32,f.U=64,f.R=128,t.exports=f},function(t,n){var r=t.exports={version:"2.4.0"};"number"==typeof __e&&(__e=r)},function(t,n,r){var e=r(26);t.exports=function(t,n,r){if(e(t),void 0===n)return t;switch(r){case 1:return function(r){return t.call(n,r)};case 2:return function(r,e){return t.call(n,r,e)};case 3:return function(r,e,i){return t.call(n,r,e,i)}}return function(){return t.apply(n,arguments)}}},function(t,n,r){var e=r(183),i=r(1),o=r(126)("metadata"),u=o.store||(o.store=new(r(186))),c=function(t,n,r){var i=u.get(t);if(!i){if(!r)return;u.set(t,i=new e)}var o=i.get(n);if(!o){if(!r)return;i.set(n,o=new e)}return o},f=function(t,n,r){var e=c(n,r,!1);return void 0!==e&&e.has(t)},a=function(t,n,r){var e=c(n,r,!1);return void 0===e?void 0:e.get(t)},s=function(t,n,r,e){c(r,e,!0).set(t,n)},l=function(t,n){var r=c(t,n,!1),e=[];return r&&r.forEach(function(t,n){e.push(n)}),e},h=function(t){return void 0===t||"symbol"==typeof t?t:String(t)},v=function(t){i(i.S,"Reflect",t)};t.exports={store:u,map:c,has:f,get:a,set:s,keys:l,key:h,exp:v}},function(t,n,r){"use strict";if(r(10)){var e=r(69),i=r(3),o=r(4),u=r(1),c=r(127),f=r(152),a=r
vendor: 4,325 bytes, line 601
601(53),s=r(68),l=r(66),h=r(27),v=r(73),p=r(67),d=r(16),y=r(75),g=r(50),b=r(24),m=r(180),x=r(114),w=r(6),S=r(17),_=r(137),O=r(70),E=r(32),P=r(71).f,j=r(154),F=r(76),M=r(7),A=r(48),N=r(117),T=r(146),I=r(155),k=r(80),L=r(123),R=r(74),C=r(130),D=r(160),U=r(11),W=r(31),G=U.f,B=W.f,V=i.RangeError,z=i.TypeError,q=i.Uint8Array,K="ArrayBuffer",J="Shared"+K,Y="BYTES_PER_ELEMENT",H="prototype",$=Array[H],X=f.ArrayBuffer,Q=f.DataView,Z=A(0),tt=A(2),nt=A(3),rt=A(4),et=A(5),it=A(6),ot=N(!0),ut=N(!1),ct=I.values,ft=I.keys,at=I.entries,st=$.lastIndexOf,lt=$.reduce,ht=$.reduceRight,vt=$.join,pt=$.sort,dt=$.slice,yt=$.toString,gt=$.toLocaleString,bt=M("iterator"),mt=M("toStringTag"),xt=F("typed_constructor"),wt=F("def_constructor"),St=c.CONSTR,_t=c.TYPED,Ot=c.VIEW,Et="Wrong length!",Pt=A(1,function(t,n){return Tt(T(t,t[wt]),n)}),jt=o(function(){return 1===new q(new Uint16Array([1]).buffer)[0]}),Ft=!!q&&!!q[H].set&&o(function(){new q(1).set({})}),Mt=function(t,n){if(void 0===t)throw z(Et);var r=+t,e=d(t);if(n&&!m(r,e))throw V(Et);return e},At=function(t,n){var r=p(t);if(r<0||r%n)throw V("Wrong offset!");return r},Nt=function(t){if(w(t)&&_t in t)return t;throw z(t+" is not a typed array!")},Tt=function(t,n){if(!(w(t)&&xt in t))throw z("It is not a typed array constructor!");return new t(n)},It=function(t,n){return kt(T(t,t[wt]),n)},kt=function(t,n){for(var r=0,e=n.length,i=Tt(t,e);e>r;)i[r]=n[r++];return i},Lt=function(t,n,r){G(t,n,{get:function(){return this._d[r]}})},Rt=function(t){var n,r,e,i,o,u,c=S(t),f=arguments.length,s=f>1?arguments[1]:void 0,l=void 0!==s,h=j(c);if(void 0!=h&&!_(h)){for(u=h.call(c),e=[],n=0;!(o=u.next()).done;n++)e.push(o.value);c=e}for(l&&f>2&&(s=a(s,arguments[2],2)),n=0,r=d(c.length),i=Tt(this,r);r>n;n++)i[n]=l?s(c[n],n):c[n];return i},Ct=function(){for(var t=0,n=arguments.length,r=Tt(this,n);n>t;)r[t]=arguments[t++];return r},Dt=!!q&&o(function(){gt.call(new q(1))}),Ut=function(){return gt.apply(Dt?dt.call(Nt(this)):Nt(this),arguments)},Wt={copyWithin:function(t,n){return D.call(Nt(this),t,n,arguments.length>2?arguments[2]:void 0)},every:function(t){return rt(Nt(this),t,arguments.length>1?arguments[1]:void 0)},fill:function(t){return C.apply(Nt(this),arguments)},filter:function(t){return It(this,tt(Nt(this),t,arguments.length>1?arguments[1]:void 0))},find:function(t){return et(Nt(this),t,arguments.length>1?arguments[1]:void 0)},findIndex:function(t){return it(Nt(this),t,arguments.length>1?arguments[1]:void 0)},forEach:function(t){Z(Nt(this),t,arguments.length>1?arguments[1]:void 0)},indexOf:function(t){return ut(Nt(this),t,arguments.length>1?arguments[1]:void 0)},includes:function(t){return ot(Nt(this),t,arguments.length>1?arguments[1]:void 0)},join:function(t){return vt.apply(Nt(this),arguments)},lastIndexOf:function(t){return st.apply(Nt(this),arguments)},map:function(t){return Pt(Nt(this),t,arguments.length>1?arguments[1]:void 0)},reduce:function(t){return lt.apply(Nt(this),arguments)},reduceRight:function(t){return ht.apply(Nt(this),arguments)},reverse:function(){for(var t,n=this,r=Nt(n).length,e=Math.floor(r/2),i=0;i<e;)t=n[i],n[i++]=n[--r],n[r]=t;return n},some:function(t){return nt(Nt(this),t,arguments.length>1?arguments[1]:void 0)},sort:function(t){return pt.call(Nt(this),t)},subarray:function(t,n){var r=Nt(this),e=r.length,i=y(t,e);return new(T(r,r[wt]))(r.buffer,r.byteOffset+i*r.BYTES_PER_ELEMENT,d((void 0===n?e:y(n,e))-i))}},Gt=function(t,n){return It(this,dt.call(Nt(this),t,n))},Bt=function(t){Nt(this);var n=At(arguments[1],1),r=this.length,e=S(t),i=d(e.length),o=0;if(i+n>r)throw V(Et);for(;o<i;)this[n+o]=e[o++]},Vt={entries:function(){return at.call(Nt(this))},keys:function(){return ft.call(Nt(this))},values:function(){return ct.call(Nt(this))}},zt=function(t,n){return w(t)&&t[_t]&&"symbol"!=typeof n&&n in t&&String(+n)==String(n)},qt=function(t,n){return zt(t,n=g(n,!0))?l(2,t[n]):B(t,n)},Kt=function(t,n,r){return!(zt(t,n=g(n,!0))&&w(r)&&b(r,"value"))||b(r,"get")||b(r,"set")||r.configurable||b(r,"writable")&&!r.writable||b(r,"enumerable")&&!r.enumerable?G(t,n,r):(t[n]=r.value,t)};St||(W.f=qt,U.f=Kt),u(u.S+u.F*!St,"Object",{getOwnPropertyDescriptor:qt,defineProperty:Kt}),o(function(){yt.call({})})&&(yt=gt=function(){return vt.call(this)});var Jt=v({},Wt);v(Jt,Vt),h(Jt,bt,Vt.values),v(Jt,{slice:Gt,set:Bt,constructor:function(){}
601,toString:yt,toLocaleString:Ut}),Lt(Jt,"buffer","b"),Lt(Jt,"byteOffset","o"),Lt(Jt,"byteLength","l"),Lt(Jt,"length","e"),G(Jt,mt,{get:function(){return this[_t]}}),t.exports=function(t,n,r,f){f=!!f;var a=t+(f?"Clamped":"")+"Array",l="Uint8Array"!=a,v="get"+t,p="set"+t,y=i[a],g=y||{},b=y&&E(y),m=!y||!c.ABV,S={},_=y&&y[H],j=function(t,r){var e=t._d;return e.v[v](r*n+e.o,jt)},F=function(t,r,e){var i=t._d;f&&(e=(e=Math.round(e))<0?0:e>255?255:255&e),i.v[p](r*n+i.o,e,jt)},M=function(t,n){G(t,n,{get:function(){return j(this,n)},set:function(t){return F(this,n,t)},enumerable:!0})};m?(y=r(function(t,r,e,i){s(t,y,a,"_d");var o,u,c,f,l=0,v=0;if(w(r)){if(!(r instanceof X||(f=x(r))==K||f==J))return _t in r?kt(y,r):Rt.call(y,r);o=r,v=At(e,n);var p=r.byteLength;if(void 0===i){if(p%n)throw V(Et);if((u=p-v)<0)throw V(Et)}else if((u=d(i)*n)+v>p)throw V(Et);c=u/n}else c=Mt(r,!0),u=c*n,o=new X(u);for(h(t,"_d",{b:o,o:v,l:u,e:c,v:new Q(o)});l<c;)M(t,l++)}),_=y[H]=O(Jt),h(_,"constructor",y)):L(function(t){new y(null),new y(t)},!0)||(y=r(function(t,r,e,i){s(t,y,a);var o;return w(r)?r instanceof X||(o=x(r))==K||o==J?void 0!==i?new g(r,At(e,n),i):void 0!==e?new g(r,At(e,n)):new g(r):_t in r?kt(y,r):Rt.call(y,r):new g(Mt(r,l))}),Z(b!==Function.prototype?P(g).concat(P(b)):P(g),function(t){t in y||h(y,t,g[t])}),y[H]=_,e||(_.constructor=y));var A=_[bt],N=!!A&&("values"==A.name||void 0==A.name),T=Vt.values;h(y,xt,!0),h(_,_t,a),h(_,Ot,!0),h(_,wt,y),(f?new y(1)[mt]==a:mt in _)||G(_,mt,{get:function(){return a}}),S[a]=y,u(u.G+u.W+u.F*(y!=g),S),u(u.S,a,{BYTES_PER_ELEMENT:n,from:Rt,of:Ct}),Y in _||h(_,Y,n),u(u.P,a,Wt),R(a),u(u.P+u.F*Ft,a,{set:Bt}),u(u.P+u.F*!N,a,Vt),u(u.P+u.F*(_.toString!=yt),a,{toString:yt}),u(u.P+u.F*o(function(){new y(1).slice()}),a,{slice:Gt}),u(u.P+u.F*(o(function(){return[1,2].toLocaleString()!=new y([1,2]).toLocaleString()})||!o(function(){_.toLocaleString.call([1,2])})),a,{toLocaleString:Ut}),k[a]=N?A:T,e||N||h(_,bt,T)}}else t.exports=function(){}},function(t,n){var r={}.toString;t.exports=function(t){return r.call(t).slice(8,-1)}},function(t,n,r){var e=r(21),i=r(5).document,o=e(i)&&e(i.createElement);t.exports=function(t){return o?i.createElement(t):{}}},function(t,n,r){t.exports=!r(12)&&!r(18)(function(){return 7!=Object.defineProperty(r(57)("div"),"a",{get:function(){return 7}}).a})},function(t,n,r){"use strict";var e=r(36),i=r(51),o=r(64),u=r(13),c=r(8),f=r(35),a=r(96),s=r(38),l=r(103),h=r(15)("iterator"),v=!([].keys&&"next"in[].keys()),p="keys",d="values",y=function(){return this};t.exports=function(t,n,r,g,b,m,x){a(r,n,g);var w,S,_,O=function(t){if(!v&&t in F)return F[t];switch(t){case p:case d:return function(){return new r(this,t)}}return function(){return new r(this,t)}},E=n+" Iterator",P=b==d,j=!1,F=t.prototype,M=F[h]||F["@@iterator"]||b&&F[b],A=M||O(b),N=b?P?O("entries"):A:void 0,T="Array"==n?F.entries||M:M;if(T&&(_=l(T.call(new t)))!==Object.prototype&&(s(_,E,!0),e||c(_,h)||u(_,h,y)),P&&M&&M.name!==d&&(j=!0,A=function(){return M.call(this)}),e&&!x||!v&&!j&&F[h]||u(F,h,A),f[n]=A,f[E]=y,b)if(w={values:P?A:O(d),keys:m?A:O(p),entries:N},x)for(S in w)S in F||o(F,S,w[S]);else i(i.P+i.F*(v||j),n,w);return w}},function(t,n,r){var e=r(20),i=r(100),o=r(34),u=r(39)("IE_PROTO"),c=function(){},f="prototype",a=function(){var t,n=r(57)("iframe"),e=o.length;for(n.style.display="none",r(93).appendChild(n),n.src="javascript:",t=n.contentWindow.document,t.open(),t.write("
601<script>document.F=Object<\/script>"),t.close(),a=t.F;e--;)delete a[f][o[e]];return a()};t.exports=Object.create||function(t,n){var r;return null!==t?(c[f]=e(t),r=new c,c[f]=null,r[u]=t):r=a(),void 0===n?r:i(r,n)}},function(t,n,r){var e=r(63),i=r(34).concat("length","prototype");n.f=Object.getOwnPropertyNames||function(t){return e(t,i)}},function(t,n){n.f=Object.getOwnPropertySymbols},function(t,n,r){var e=r(8),i=r(9),o=r(90)(!1),u=r(39)("IE_PROTO");t.exports=function(t,n){var r,c=i(t),f=0,a=[];for(r in c)r!=u&&e(c,r)&&a.push(r);for(;n.length>f;)e(c,r=n[f++])&&(~o(a,r)||a.push(r));return a}},function(t,n,r){t.exports=r(13)},function(t,n,r){var e=r(76)("meta"),i=r(6),o=r(24),u=r(11).f,c=0,f=Object.isExtensible||function(){return!0},a=!r(4)(function(){return f(Object.preventExtensions({}))}),s=function(t){u(t,e,{value:{i:"O"+ ++c,w:{}}})},l=function(t,n){if(!i(t))return"symbol"==typeof t?t:("string"==typeof t?"S":"P")+t;if(!o(t,e)){if(!f(t))return"F";if(!n)return"E";s(t)}return t[e].i},h=function(t,n){if(!o(t,e)){if(!f(t))return!0;if(!n)return!1;s(t)}return t[e].w},v=function(t){return a&&p.NEED&&f(t)&&!o(t,e)&&s(t),t},p=t.exports={KEY:e,NEED:!1,fastKey:l,getWeak:h,onFreeze:v}},function(t,n){t.exports=function(t,n){return{enumerable:!(1&t),configurable:!(2&t),writable:!(4&t),value:n}}},function(t,n){var r=Math.ceil,e=Math.floor;t.exports=function(t){return isNaN(t=+t)?0:(t>0?e:r)(t)}},function(t,n){t.exports=function(t,n,r,e){if(!(t instanceof n)||void 0!==e&&e in t)throw TypeError(r+": incorrect invocation!");return t}},function(t,n){t.exports=!1},function(t,n,r){var e=r(2),i=r(173),o=r(133),u=r(145)("IE_PROTO"),c=function(){},f="prototype",a=function(){var t,n=r(132)("iframe"),e=o.length;for(n.style.display="none",r(135).appendChild(n),n.src="javascript:",t=n.contentWindow.document,t.open(),t.write("
601<script>document.F=Object<\/script>"),t.close(),a=t.F;e--;)delete a[f][o[e]];return a()};t.exports=Object.create||function(t,n){var r;return null!==t?(c[f]=e(t),r=new c,c[f]=null,r[u]=t):r=a(),void 0===n?r:i(r,n)}},function(t,n,r){var e=r(175),i=r(133).concat("length","prototype");n.f=Object.getOwnPropertyNames||function(t){return e(t,i)}},function(t,n,r){var e=r(175),i=r(133);t.exports=Object.keys||function(t){return e(t,i)}},function(t,n,r){var e=r(28);t.exports=function(t,n,r){for(var i in n)e(t,i,n[i],r);return t}},function(t,n,r){"use strict";var e=r(3),i=r(11),o=r(10),u=r(7)("species");t.exports=function(t){var n=e[t];o&&n&&!n[u]&&i.f(n,u,{configurable:!0,get:function(){return this}})}},function(t,n,r){var e=r(67),i=Math.max,o=Math.min;t.exports=function(t,n){return t=e(t),t<0?i(t+n,0):o(t,n)}},function(t,n){var r=0,e=Math.random();t.exports=function(t){return"Symbol(".concat(void 0===t?"":t,")_",(++r+e).toString(36))}},function(t,n,r){var e=r(33);t.exports=function(t){return Object(e(t))}},function(t,n,r){var e=r(7)("unscopables"),i=Array.prototype;void 0==i[e]&&r(27)(i,e,{}),t.exports=function(t){i[e][t]=!0}},function(t,n,r){var e=r(53),i=r(169),o=r(137),u=r(2),c=r(16),f=r(154),a={},s={},n=t.exports=function(t,n,r,l,h){var v,p,d,y,g=h?function(){return t}:f(t),b=e(r,l,n?2:1),m=0;if("function"!=typeof g)throw TypeError(t+" is not iterable!");if(o(g)){for(v=c(t.length);v>m;m++)if((y=n?b(u(p=t[m])[0],p[1]):b(t[m]))===a||y===s)return y}else for(d=g.call(t);!(p=d.next()).done;)if((y=i(d,b,p.value,n))===a||y===s)return y};n.BREAK=a,n.RETURN=s},function(t,n){t.exports={}},function(t,n,r){var e=r(11).f,i=r(24),o=r(7)("toStringTag");t.exports=function(t,n,r){t&&!i(t=r?t:t.prototype,o)&&e(t,o,{configurable:!0,value:n})}},function(t,n,r){var e=r(1),i=r(46),o=r(4),u=r(150),c="["+u+"]",f="​…",a=RegExp("^"+c+c+"*"),s=RegExp(c+c+"*$"),l=function(t,n,r){var i={},c=o(function(){return!!u[t]()||f[t]()!=f}),a=i[t]=c?n(h):u[t];r&&(i[r]=a),e(e.P+e.F*c,"String",i)},h=l.trim=function(t,n){return t=String(i(t)),1&n&&(t=t.replace(a,"")),2&n&&(t=t.replace(s,"")),t};t.exports=l},function(t,n,r){t.exports={default:r(86),__esModule:!0}},function(t,n,r){t.exports={default:r(87),__esModule:!0}},function(t,n,r){"use strict";function e(t){return t&&t.__esModule?t:{default:t}}n.__esModule=!0;var i=r(84),o=e(i),u=r(83),c=e(u),f="function"==typeof c.default&&"symbol"==typeof o.default?function(t){return typeof t}:function(t){return t&&"function"==typeof c.default&&t.constructor===c.default&&t!==c.default.prototype?"symbol":typeof t};n.default="function"==typeof c.default&&"symbol"===f(o.default)?function(t){return void 0===t?"undefined":f(t)}:function(t){return t&&"function"==typeof c.default&&t.constructor===c.default&&t!==c.default.prototype?"symbol":void 0===t?"undefined":f(t)}},function(t,n,r){r(110),r(108),r(111),r(112),t.exports=r(25).Symbol},function(t,n,r){r(109),r(113),t.exports=r(44).f("iterator")},function(t,n){t.exports=function(t){if("function"!=typeof t)throw TypeError(t+" is not a function!");return t}},function(t,n){t.exports=function(){}},function(t,n,r){var e=r(9),i=r(106),o=r(105);t.exports=function(t){return function(n,r,u){var c,f=e(n),a=i(f.length),s=o(u,a);if(t&&r!=r){for(;a>s;)if((c=f[s++])!=c)return!0}else for(;a>s;s++)if((t||s in f)&&f[s]===r)return t||s||0;return!t&&-1}}},function(t,n,r){var e=r(88);t.exports=function(t,n,r){if(e(t),void 0===n)return t;switch(r){case 1:return function(r){return t.call(n,r)};case 2:return function(r,e){return t.call(n,r,e)};case 3:return function(r,e,i){return t.call(n,r,e,i)}}return function(){return t.apply(n,arguments)}}},function(t,n,r){var e=r(19),i=r(62),o=r(37);t.exports=function(t){var n=e(t),r=i.f;if(r)for(var u,c=r(t),f=o.f,a=0;c.length>a;)f.call(t,u=c[a++])&&n.push(u);return n}},function(t,n,r){t.exports=r(5).document&&document.documentElement},function(t,n,r){var e=r(56);t.exports=Object("z").propertyIsEnumerable(0)?Object:function(t){return"String"==e(t)?t.split(""):Object(t)}},function(t,n,r){var e=r(56);t.exports=Array.isArray||function(t){return"Array"==e(t)}},function(t,n,r){"use strict";var e=r(60),i=r(22),o=r(38),u={};r(13)(u,r(15)("iterator"),function(){return this}),t.exports=function(t,n,r){t.prototype=e(u,{next:i(1,r)}),o(t,n+" Iterator")}},function(t,n){t.exports=function(t,n){return{value:n,done:!!t}}},function(t,n,r){var e=r(19),i=r(9);t.exports=function(t,n){for(var r,o=i(t),u=e(o),c=u.length,f=0;c>f;)if(o[r=u[f++]]===n)return r}},function(t,n,r){var e=r(23)("meta"),i=r(21),o=r(8),u=r(14).f,c=0,f=Object.isExtensible||function(){return!0},a=!r(18)(function(){return f(Object.preventExtensions({}))}),s=function(t){u(t,e,{value:{i:"O"+ ++c,w:{}}
vendor: 7,167 bytes, line 601
601})},l=function(t,n){if(!i(t))return"symbol"==typeof t?t:("string"==typeof t?"S":"P")+t;if(!o(t,e)){if(!f(t))return"F";if(!n)return"E";s(t)}return t[e].i},h=function(t,n){if(!o(t,e)){if(!f(t))return!0;if(!n)return!1;s(t)}return t[e].w},v=function(t){return a&&p.NEED&&f(t)&&!o(t,e)&&s(t),t},p=t.exports={KEY:e,NEED:!1,fastKey:l,getWeak:h,onFreeze:v}},function(t,n,r){var e=r(14),i=r(20),o=r(19);t.exports=r(12)?Object.defineProperties:function(t,n){i(t);for(var r,u=o(n),c=u.length,f=0;c>f;)e.f(t,r=u[f++],n[r]);return t}},function(t,n,r){var e=r(37),i=r(22),o=r(9),u=r(42),c=r(8),f=r(58),a=Object.getOwnPropertyDescriptor;n.f=r(12)?a:function(t,n){if(t=o(t),n=u(n,!0),f)try{return a(t,n)}catch(t){}if(c(t,n))return i(!e.f.call(t,n),t[n])}},function(t,n,r){var e=r(9),i=r(61).f,o={}.toString,u="object"==typeof window&&window&&Object.getOwnPropertyNames?Object.getOwnPropertyNames(window):[],c=function(t){try{return i(t)}catch(t){return u.slice()}};t.exports.f=function(t){return u&&"[object Window]"==o.call(t)?c(t):i(e(t))}},function(t,n,r){var e=r(8),i=r(77),o=r(39)("IE_PROTO"),u=Object.prototype;t.exports=Object.getPrototypeOf||function(t){return t=i(t),e(t,o)?t[o]:"function"==typeof t.constructor&&t instanceof t.constructor?t.constructor.prototype:t instanceof Object?u:null}},function(t,n,r){var e=r(41),i=r(33);t.exports=function(t){return function(n,r){var o,u,c=String(i(n)),f=e(r),a=c.length;return f<0||f>=a?t?"":void 0:(o=c.charCodeAt(f),o<55296||o>56319||f+1===a||(u=c.charCodeAt(f+1))<56320||u>57343?t?c.charAt(f):o:t?c.slice(f,f+2):u-56320+(o-55296<<10)+65536)}}},function(t,n,r){var e=r(41),i=Math.max,o=Math.min;t.exports=function(t,n){return t=e(t),t<0?i(t+n,0):o(t,n)}},function(t,n,r){var e=r(41),i=Math.min;t.exports=function(t){return t>0?i(e(t),9007199254740991):0}},function(t,n,r){"use strict";var e=r(89),i=r(97),o=r(35),u=r(9);t.exports=r(59)(Array,"Array",function(t,n){this._t=u(t),this._i=0,this._k=n},function(){var t=this._t,n=this._k,r=this._i++;return!t||r>=t.length?(this._t=void 0,i(1)):"keys"==n?i(0,r):"values"==n?i(0,t[r]):i(0,[r,t[r]])},"values"),o.Arguments=o.Array,e("keys"),e("values"),e("entries")},function(t,n){},function(t,n,r){"use strict";var e=r(104)(!0);r(59)(String,"String",function(t){this._t=String(t),this._i=0},function(){var t,n=this._t,r=this._i;return r>=n.length?{value:void 0,done:!0}:(t=e(n,r),this._i+=t.length,{value:t,done:!1})})},function(t,n,r){"use strict";var e=r(5),i=r(8),o=r(12),u=r(51),c=r(64),f=r(99).KEY,a=r(18),s=r(40),l=r(38),h=r(23),v=r(15),p=r(44),d=r(43),y=r(98),g=r(92),b=r(95),m=r(20),x=r(9),w=r(42),S=r(22),_=r(60),O=r(102),E=r(101),P=r(14),j=r(19),F=E.f,M=P.f,A=O.f,N=e.Symbol,T=e.JSON,I=T&&T.stringify,k="prototype",L=v("_hidden"),R=v("toPrimitive"),C={}.propertyIsEnumerable,D=s("symbol-registry"),U=s("symbols"),W=s("op-symbols"),G=Object[k],B="function"==typeof N,V=e.QObject,z=!V||!V[k]||!V[k].findChild,q=o&&a(function(){return 7!=_(M({},"a",{get:function(){return M(this,"a",{value:7}).a}})).a})?function(t,n,r){var e=F(G,n);e&&delete G[n],M(t,n,r),e&&t!==G&&M(G,n,e)}:M,K=function(t){var n=U[t]=_(N[k]);return n._k=t,n},J=B&&"symbol"==typeof N.iterator?function(t){return"symbol"==typeof t}:function(t){return t instanceof N},Y=function(t,n,r){return t===G&&Y(W,n,r),m(t),n=w(n,!0),m(r),i(U,n)?(r.enumerable?(i(t,L)&&t[L][n]&&(t[L][n]=!1),r=_(r,{enumerable:S(0,!1)})):(i(t,L)||M(t,L,S(1,{})),t[L][n]=!0),q(t,n,r)):M(t,n,r)},H=function(t,n){m(t);for(var r,e=g(n=x(n)),i=0,o=e.length;o>i;)Y(t,r=e[i++],n[r]);return t},$=function(t,n){return void 0===n?_(t):H(_(t),n)},X=function(t){var n=C.call(this,t=w(t,!0));return!(this===G&&i(U,t)&&!i(W,t))&&(!(n||!i(this,t)||!i(U,t)||i(this,L)&&this[L][t])||n)},Q=function(t,n){if(t=x(t),n=w(n,!0),t!==G||!i(U,n)||i(W,n)){var r=F(t,n);return!r||!i(U,n)||i(t,L)&&t[L][n]||(r.enumerable=!0),r}},Z=function(t){for(var n,r=A(x(t)),e=[],o=0;r.length>o;)i(U,n=r[o++])||n==L||n==f||e.push(n);return e},tt=function(t){for(var n,r=t===G,e=A(r?W:x(t)),o=[],u=0;e.length>u;)!i(U,n=e[u++])||r&&!i(G,n)||o.push(U[n]);return o};B||(N=function(){if(this instanceof N)throw TypeError("Symbol is not a constructor!");var t=h(arguments.length>0?arguments[0]:void 0),n=function(r){this===G&&n.call(W,r),i(this,L)&&i(this[L],t)&&(this[L][t]=!1),q(this,t,S(1,r))};return o&&z&&q(G,t,{configurable:!0,set:n}),K(t)},c(N[k],"toString",function(){return this._k}),E.f=Q,P.f=Y,r(61).f=O.f=Z,r(37).f=X,r(62).f=tt,o&&!r(36)&&c(G,"propertyIsEnumerable",X,!0),p.f=function(t){return K(v(t))}),u(u.G+u.W+u.F*!B,{Symbol:N});for(var nt="hasInstance,isConcatSpreadable,iterator,match,replace,search,species,split,toPrimitive,toStringTag,unscopables".split(","),rt=0;nt.length>rt;)v(nt[rt++]);for(var nt=j(v.store),rt=0;nt.length>rt;)d(nt[rt++]);u(u.S+u.F*!B,"Symbol",{for:function(t){return i(D,t+="")?D[t]:D[t]=N(t)},keyFor:function(t){if(J(t))return y(D,t);throw TypeError(t+" is not a symbol!")},useSetter:function(){z=!0},useSimple:function(){z=!1}}),u(u.S+u.F*!B,"Object",{create:$,defineProperty:Y,defineProperties:H,getOwnPropertyDescriptor:Q,getOwnPropertyNames:Z,getOwnPropertySymbols:tt}),T&&u(u.S+u.F*(!B||a(function(){var t=N();return"[null]"!=I([t])||"{}"!=I({a:t})||"{}"!=I(Object(t))})),"JSON",{stringify:function(t){if(void 0!==t&&!J(t)){for(var n,r,e=[t],i=1;arguments.length>i;)e.push(arguments[i++]);return n=e[1],"function"==typeof n&&(r=n),!r&&b(n)||(n=function(t,n){if(r&&(n=r.call(this,t,n)),!J(n))return n}),e[1]=n,I.apply(T,e)}}}),N[k][R]||r(13)(N[k],R,N[k].valueOf),l(N,"Symbol"),l(Math,"Math",!0),l(e.JSON,"JSON",!0)},function(t,n,r){r(43)("asyncIterator")},function(t,n,r){r(43)("observable")},function(t,n,r){r(107);for(var e=r(5),i=r(13),o=r(35),u=r(15)("toStringTag"),c=["NodeList","DOMTokenList","MediaList","StyleSheetList","CSSRuleList"],f=0;f<5;f++){var a=c[f],s=e[a],l=s&&s.prototype;l&&!l[u]&&i(l,u,a),o[a]=o.Array}},function(t,n,r){var e=r(45),i=r(7)("toStringTag"),o="Arguments"==e(function(){return arguments}()),u=function(t,n){try{return t[n]}catch(t){}};t.exports=function(t){var n,r,c;return void 0===t?"Undefined":null===t?"Null":"string"==typeof(r=u(n=Object(t),i))?r:o?e(n):"Object"==(c=e(n))&&"function"==typeof n.callee?"Arguments":c}},function(t,n,r){var e=r(45);t.exports=Object("z").propertyIsEnumerable(0)?Object:function(t){return"String"==e(t)?t.split(""):Object(t)}},function(t,n){n.f={}.propertyIsEnumerable},function(t,n,r){var e=r(30),i=r(16),o=r(75);t.exports=function(t){return function(n,r,u){var c,f=e(n),a=i(f.length),s=o(u,a);if(t&&r!=r){for(;a>s;)if((c=f[s++])!=c)return!0}else for(;a>s;s++)if((t||s in f)&&f[s]===r)return t||s||0;return!t&&-1}}},function(t,n,r){"use strict";var e=r(3),i=r(1),o=r(28),u=r(73),c=r(65),f=r(79),a=r(68),s=r(6),l=r(4),h=r(123),v=r(81),p=r(136);t.exports=function(t,n,r,d,y,g){var b=e[t],m=b,x=y?"set":"add",w=m&&m.prototype,S={},_=function(t){var n=w[t];o(w,t,"delete"==t?function(t){return!(g&&!s(t))&&n.call(this,0===t?0:t)}:"has"==t?function(t){return!(g&&!s(t))&&n.call(this,0===t?0:t)}:"get"==t?function(t){return g&&!s(t)?void 0:n.call(this,0===t?0:t)}:"add"==t?function(t){return n.call(this,0===t?0:t),this}:function(t,r){return n.call(this,0===t?0:t,r),this})};
601if("function"==typeof m&&(g||w.forEach&&!l(function(){(new m).entries().next()}))){var O=new m,E=O[x](g?{}:-0,1)!=O,P=l(function(){O.has(1)}),j=h(function(t){new m(t)}),F=!g&&l(function(){for(var t=new m,n=5;n--;)t[x](n,n);return!t.has(-0)});j||(m=n(function(n,r){a(n,m,t);var e=p(new b,n,m);return void 0!=r&&f(r,y,e[x],e),e}),m.prototype=w,w.constructor=m),(P||F)&&(_("delete"),_("has"),y&&_("get")),(F||E)&&_(x),g&&w.clear&&delete w.clear}else m=d.getConstructor(n,t,y,x),u(m.prototype,r),c.NEED=!0;return v(m,t),S[t]=m,i(i.G+i.W+i.F*(m!=b),S),g||d.setStrong(m,t,y),m}},function(t,n,r){"use strict";var e=r(27),i=r(28),o=r(4),u=r(46),c=r(7);t.exports=function(t,n,r){var f=c(t),a=r(u,f,""[t]),s=a[0],l=a[1];o(function(){var n={};return n[f]=function(){return 7},7!=""[t](n)})&&(i(String.prototype,t,s),e(RegExp.prototype,f,2==n?function(t,n){return l.call(t,this,n)}:function(t){return l.call(t,this)}))}
602},function(t,n,r){"use strict";var e=r(2);t.exports=function(){var t=e(this),n="";return t.global&&(n+="g"),t.ignoreCase&&(n+="i"),t.multiline&&(n+="m"),t.unicode&&(n+="u"),t.sticky&&(n+="y"),n}},function(t,n){t.exports=function(t,n,r){var e=void 0===r;switch(n.length){case 0:return e?t():t.call(r);case 1:return e?t(n[0]):t.call(r,n[0]);case 2:return e?t(n[0],n[1]):t.call(r,n[0],n[1]);case 3:return e?t(n[0],n[1],n[2]):t.call(r,n[0],n[1],n[2]);case 4:return e?t(n[0],n[1],n[2],n[3]):t.call(r,n[0],n[1],n[2],n[3])}return t.apply(r,n)}},function(t,n,r){var e=r(6),i=r(45),o=r(7)("match");t.exports=function(t){var n;return e(t)&&(void 0!==(n=t[o])?!!n:"RegExp"==i(t))}},function(t,n,r){var e=r(7)("iterator"),i=!1;try{var o=[7][e]();o.return=function(){i=!0},Array.from(o,function(){throw 2})}catch(t){}t.exports=function(t,n){if(!n&&!i)return!1;var r=!1;try{var o=[7],u=o[e]();u.next=function(){return{done:r=!0}},o[e]=function(){return u},t(o)}catch(t){}return r}},function(t,n,r){t.exports=r(69)||!r(4)(function(){var t=Math.random();__defineSetter__.call(null,t,function(){}),delete r(3)[t]})},function(t,n){n.f=Object.getOwnPropertySymbols},function(t,n,r){var e=r(3),i="__core-js_shared__",o=e[i]||(e[i]={});t.exports=function(t){return o[t]||(o[t]={})}},function(t,n,r){for(var e,i=r(3),o=r(27),u=r(76),c=u("typed_array"),f=u("view"),a=!(!i.ArrayBuffer||!i.DataView),s=a,l=0,h="Int8Array,Uint8Array,Uint8ClampedArray,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array".split(",");l<9;)(e=i[h[l++]])?(o(e.prototype,c,!0),o(e.prototype,f,!0)):s=!1;t.exports={ABV:a,CONSTR:s,TYPED:c,VIEW:f}},function(t,n){"use strict";var r={versions:function(){var t=window.navigator.userAgent;return{trident:t.indexOf("Trident")>-1,presto:t.indexOf("Presto")>-1,webKit:t.indexOf("AppleWebKit")>-1,gecko:t.indexOf("Gecko")>-1&&-1==t.indexOf("KHTML"),mobile:!!t.match(/AppleWebKit.*Mobile.*/),ios:!!t.match(/\(i[^;]+;( U;)? CPU.+Mac OS X/),android:t.indexOf("Android")>-1||t.indexOf("Linux")>-1,iPhone:t.indexOf("iPhone")>-1||t.indexOf("Mac")>-1,iPad:t.indexOf("iPad")>-1,webApp:-1==t.indexOf("Safari"),weixin:-1==t.indexOf("MicroMessenger")}}()};t.exports=r},function(t,n,r){"use strict";var e=r(85),i=function(t){return t&&t.__esModule?t:{default:t}}(e),o=function(){function t(t,n,e){return n||e?String.fromCharCode(n||e):r[t]||t}function n(t){return e[t]}var r={"&quot;":'"',"&lt;":"<","&gt;":">","&amp;":"&","&nbsp;":" "},e={};for(var u in r)e[r[u]]=u;return r["&apos;"]="'",e["'"]="&#39;",{encode:function(t){return t?(""+t).replace(/['<> "&]/g,n).replace(/\r?\n/g,"<br/>").replace(/\s/g,"&nbsp;"):""},decode:function(n){return n?(""+n).replace(/<br\s*\/?>/gi,"\n").replace(/&quot;|&lt;|&gt;|&amp;|&nbsp;|&apos;|&#(\d+);|&#(\d+)/g,t).replace(/\u00a0/g," "):""},encodeBase16:function(t){if(!t)return t;t+="";for(var n=[],r=0,e=t.length;e>r;r++)n.push(t.charCodeAt(r).toString(16).toUpperCase());return n.join("")},encodeBase16forJSON:function(t){if(!t)return t;t=t.replace(/[\u4E00-\u9FBF]/gi,function(t){return escape(t).replace("%u","\\u")});for(var n=[],r=0,e=t.length;e>r;r++)n.push(t.charCodeAt(r).toString(16).toUpperCase());return n.join("")},decodeBase16:function(t){if(!t)return t;t+="";for(var n=[],r=0,e=t.length;e>r;r+=2)n.push(String.fromCharCode("0x"+t.slice(r,r+2)));return n.join("")},encodeObject:function(t){if(t instanceof Array)for(var n=0,r=t.length;r>n;n++)t[n]=o.encodeObject(t[n]);else if("object"==(void 0===t?"undefined":(0,i.default)(t)))for(var e in t)t[e]=o.encodeObject(t[e]);else if("string"==typeof t)return o.encode(t);return t},loadScript:function(t){var n=document.createElement("script");document.getElementsByTagName("body")[0].appendChild(n),n.setAttribute("src",t)},addLoadEvent:function(t){var n=window.onload;"function"!=typeof window.onload?window.onload=t:window.onload=function(){n(),t()}}}}();t.exports=o},function(t,n,r){"use strict";var e=r(17),i=r(75),o=r(16);t.exports=function(t){for(var n=e(this),r=o(n.length),u=arguments.length,c=i(u>1?arguments[1]:void 0,r),f=u>2?arguments[2]:void 0,a=void 0===f?r:i(f,r);a>c;)n[c++]=t;return n}},function(t,n,r){"use strict";var e=r(11),i=r(66);t.exports=function(t,n,r){n in t?e.f(t,n,i(0,r)):t[n]=r}},function(t,n,r){var e=r(6),i=r(3).document,o=e(i)&&e(i.createElement);t.exports=function(t){return o?i.createElement(t):{}}},function(t,n){t.exports="constructor,hasOwnProperty,isPrototypeOf,propertyIsEnumerable,toLocaleStr
602ing,toString,valueOf".split(",")},function(t,n,r){var e=r(7)("match");t.exports=function(t){var n=/./;try{"/./"[t](n)}catch(r){try{return n[e]=!1,!"/./"[t](n)}catch(t){}}return!0}},function(t,n,r){t.exports=r(3).document&&document.documentElement},function(t,n,r){var e=r(6),i=r(144).set;t.exports=function(t,n,r){var o,u=n.constructor;return u!==r&&"function"==typeof u&&(o=u.prototype)!==r.prototype&&e(o)&&i&&i(t,o),t}},function(t,n,r){var e=r(80),i=r(7)("iterator"),o=Array.prototype;t.exports=function(t){return void 0!==t&&(e.Array===t||o[i]===t)}},function(t,n,r){var e=r(45);t.exports=Array.isArray||function(t){return"Array"==e(t)}},function(t,n,r){"use strict";var e=r(70),i=r(66),o=r(81),u={};r(27)(u,r(7)("iterator"),function(){return this}),t.exports=function(t,n,r){t.prototype=e(u,{next:i(1,r)}),o(t,n+" Iterator")}},function(t,n,r){"use strict";var e=r(69),i=r(1),o=r(28),u=r(27),c=r(24),f=r(80),a=r(139),s=r(81),l=r(32),h=r(7)("iterator"),v=!([].keys&&"next"in[].keys()),p="keys",d="values",y=function(){return this};t.exports=function(t,n,r,g,b,m,x){a(r,n,g);var w,S,_,O=function(t){if(!v&&t in F)return F[t];switch(t){case p:case d:return function(){return new r(this,t)}}return function(){return new r(this,t)}},E=n+" Iterator",P=b==d,j=!1,F=t.prototype,M=F[h]||F["@@iterator"]||b&&F[b],A=M||O(b),N=b?P?O("entries"):A:void 0,T="Array"==n?F.entries||M:M;if(T&&(_=l(T.call(new t)))!==Object.prototype&&(s(_,E,!0),e||c(_,h)||u(_,h,y)),P&&M&&M.name!==d&&(j=!0,A=function(){return M.call(this)}),e&&!x||!v&&!j&&F[h]||u(F,h,A),f[n]=A,f[E]=y,b)if(w={values:P?A:O(d),keys:m?A:O(p),entries:N},x)for(S in w)S in F||o(F,S,w[S]);else i(i.P+i.F*(v||j),n,w);return w}},function(t,n){var r=Math.expm1;t.exports=!r||r(10)>22025.465794806718||r(10)<22025.465794806718||-2e-17!=r(-2e-17)?function(t){return 0==(t=+t)?t:t>-1e-6&&t<1e-6?t+t*t/2:Math.exp(t)-1}:r},function(t,n){t.exports=Math.sign||function(t){return 0==(t=+t)||t!=t?t:t<0?-1:1}},function(t,n,r){var e=r(3),i=r(151).set,o=e.MutationObserver||e.WebKitMutationObserver,u=e.process,c=e.Promise,f="process"==r(45)(u);t.exports=function(){var t,n,r,a=function(){var e,i;for(f&&(e=u.domain)&&e.exit();t;){i=t.fn,t=t.next;try{i()}catch(e){throw t?r():n=void 0,e}}n=void 0,e&&e.enter()};if(f)r=function(){u.nextTick(a)};else if(o){var s=!0,l=document.createTextNode("");new o(a).observe(l,{characterData:!0}),r=function(){l.data=s=!s}}else if(c&&c.resolve){var h=c.resolve();r=function(){h.then(a)}}else r=function(){i.call(e,a)};return function(e){var i={fn:e,next:void 0};n&&(n.next=i),t||(t=i,r()),n=i}}},function(t,n,r){var e=r(6),i=r(2),o=function(t,n){if(i(t),!e(n)&&null!==n)throw TypeError(n+": can't set as prototype!")};t.exports={set:Object.setPrototypeOf||("__proto__"in{}?function(t,n,e){try{e=r(53)(Function.call,r(31).f(Object.prototype,"__proto__").set,2),e(t,[]),n=!(t instanceof Array)}catch(t){n=!0}return function(t,r){return o(t,r),n?t.__proto__=r:e(t,r),t}}({},!1):void 0),check:o}},function(t,n,r){var e=r(126)("keys"),i=r(76);t.exports=function(t){return e[t]||(e[t]=i(t))}},function(t,n,r){var e=r(2),i=r(26),o=r(7)("species");t.exports=function(t,n){var r,u=e(t).constructor;return void 0===u||void 0==(r=e(u)[o])?n:i(r)}},function(t,n,r){var e=r(67),i=r(46);t.exports=function(t){return function(n,r){var o,u,c=String(i(n)),f=e(r),a=c.length;return f<0||f>=a?t?"":void 0:(o=c.charCodeAt(f),o<55296||o>56319||f+1===a||(u=c.charCodeAt(f+1))<56320||u>57343?t?c.charAt(f):o:t?c.slice(f,f+2):u-56320+(o-55296<<10)+65536)}}},function(t,n,r){var e=r(122),i=r(46);t.exports=function(t,n,r){if(e(n))throw TypeError("String#"+r+" doesn't accept regex!");return String(i(t))}},function(t,n,r){"use strict";var e=r(67),i=r(46);t.exports=function(t){var n=String(i(this)),r="",o=e(t);if(o<0||o==1/0)throw RangeError("Count can't be negative");for(;o>0;(o>>>=1)&&(n+=n))1&o&&(r+=n);return r}},function(t,n){t.exports="\t\n\v\f\r   ᠎              \u2028\u2029\ufeff"},function(t,n,r){var e,i,o,u=r(53),c=r(121),f=r(135),a=r(132),s=r(3),l=s.process,h=s.setImmediate,v=s.clearImmediate,p=s.MessageChannel,d=0,y={},g="onreadystatechange",b=function(){var t=+this;if(y.hasOwnProperty(t)){var n=y[t];delete y[t],n()}},m=function(t){b.call(t.data)};h&&v||(h=function(t){for(var n=[],r=1;arguments.length>r;)n.push(arguments[r++]);return y[++d]=function(){c("function"==typeof t?t:Function(t),n)},e(d),d},v=function(t){delete y[t]},"process"==r(45)(l)?e=function(t){l.nextTick(u(b,t,1))}:p?(i=new p,o=i.port2,i.port1.onmessage=m,e=u(o.postMessage,o,1)):s.addEventListener&&"function"==typeof postMessage&&!s.importScripts?(e=function(t){s.postMessage(t+"","*")},s.addEventListener("message",m,!1)):e=g in a("script")?function(t){f.appendChild(a("script"))[g]=function(){f.removeChild(this),b.call(t)}}:function(t){setTimeout(u(b,t,1),0)}),t.exports={set:h,clear:v}},function(t,n,r){"use strict";var e=r(3),i=r(10),o=r(69),u=r(127),c=r(27),f=r(73),a=r(4),s=r(68),l=r(67),h=r(16),v=r(71).f,p=r(11).f,d=r(130),y=r(81),g="ArrayBuffer",b="DataView",m="prototype",x="Wrong length!",w="Wrong index!",S=e[g],_=e[b],O=e.Math,E=e.RangeError,P=e.Infinity,j=S,F=O.abs,M=O.pow,A=O.floor,N=O.log,T=O.LN2,I="buffer",k="byteLength",L="byteOffset",R=i?"_b":I,C=i?"_l":k,D=i?"_o":L,U=function(t,n,r){var e,i,o,u=Array(r),c=8*r-n-1,f=(1<<c)-1,a=f>>1,s=23===n?M(2,-24)-M(2,-77):0,l=0,h=t<0||0===t&&1/t<0?1:0;for(t=F(t),t!=t||t===P?(i=t!=t?1:0,e=f):(e=A(N(t)/T),t*(o=M(2,-e))<1&&(e--,o*=2),t+=e+a>=1?s/o:s*M(2,1-a),t*o>=2&&(e++,o/=2),e+a>=f?(i=0,e=f):e+a>=1?(i=(t*o-1)*M(2,n),e+=a):(i=t*M(2,a-1)*M(2,n),e=0));n>=8;u[l++]=255&i,i/=256,n-=8);for(e=e<<n|i,c+=n;c>0;
602u[l++]=255&e,e/=256,c-=8);return u[--l]|=128*h,u},W=function(t,n,r){var e,i=8*r-n-1,o=(1<<i)-1,u=o>>1,c=i-7,f=r-1,a=t[f--],s=127&a;for(a>>=7;c>0;s=256*s+t[f],f--,c-=8);for(e=s&(1<<-c)-1,s>>=-c,c+=n;c>0;e=256*e+t[f],f--,c-=8);if(0===s)s=1-u;else{if(s===o)return e?NaN:a?-P:P;e+=M(2,n),s-=u}return(a?-1:1)*e*M(2,s-n)},G=function(t){return t[3]<<24|t[2]<<16|t[1]<<8|t[0]},B=function(t){return[255&t]},V=function(t){return[255&t,t>>8&255]},z=function(t){return[255&t,t>>8&255,t>>16&255,t>>24&255]},q=function(t){return U(t,52,8)},K=function(t){return U(t,23,4)},J=function(t,n,r){p(t[m],n,{get:function(){return this[r]}})},Y=function(t,n,r,e){var i=+r,o=l(i);if(i!=o||o<0||o+n>t[C])throw E(w);var u=t[R]._b,c=o+t[D],f=u.slice(c,c+n);return e?f:f.reverse()},H=function(t,n,r,e,i,o){var u=+r,c=l(u);if(u!=c||c<0||c+n>t[C])throw E(w);for(var f=t[R]._b,a=c+t[D],s=e(+i),h=0;h<n;h++)f[a+h]=s[o?h:n-h-1]},$=function(t,n){s(t,S,g);var r=+n,e=h(r);if(r!=e)throw E(x);return e};if(u.ABV){if(!a(function(){new S})||!a(function(){new S(.5)})){S=function(t){return new j($(this,t))};for(var X,Q=S[m]=j[m],Z=v(j),tt=0;Z.length>tt;)(X=Z[tt++])in S||c(S,X,j[X]);o||(Q.constructor=S)}var nt=new _(new S(2)),rt=_[m].setInt8;nt.setInt8(0,2147483648),nt.setInt8(1,2147483649),!nt.getInt8(0)&&nt.getInt8(1)||f(_[m],{setInt8:function(t,n){rt.call(this,t,n<<24>>24)},setUint8:function(t,n){rt.call(this,t,n<<24>>24)}},!0)}else S=function(t){var n=$(this,t);this._b=d.call(Array(n),0),this[C]=n},_=function(t,n,r){s(this,_,b),s(t,S,b);var e=t[C],i=l(n);if(i<0||i>e)throw E("Wrong offset!");if(r=void 0===r?e-i:h(r),i+r>e)throw E(x);this[R]=t,this[D]=i,this[C]=r},i&&(J(S,k,"_l"),J(_,I,"_b"),J(_,k,"_l"),J(_,L,"_o")),f(_[m],{getInt8:function(t){return Y(this,1,t)[0]<<24>>24},getUint8:function(t){return Y(this,1,t)[0]},getInt16:function(t){var n=Y(this,2,t,arguments[1]);return(n[1]<<8|n[0])<<16>>16},getUint16:function(t){var n=Y(this,2,t,arguments[1]);return n[1]<<8|n[0]},getInt32:function(t){return G(Y(this,4,t,arguments[1]))},getUint32:function(t){return G(Y(this,4,t,arguments[1]))>>>0},getFloat32:function(t){return W(Y(this,4,t,arguments[1]),23,4)},getFloat64:function(t){return W(Y(this,8,t,arguments[1]),52,8)},setInt8:function(t,n){H(this,1,t,B,n)},setUint8:function(t,n){H(this,1,t,B,n)},setInt16:function(t,n){H(this,2,t,V,n,arguments[2])},setUint16:function(t,n){H(this,2,t,V,n,arguments[2])},setInt32:function(t,n){H(this,4,t,z,n,arguments[2])},setUint32:function(t,n){H(this,4,t,z,n,arguments[2])},setFloat32:function(t,n){H(this,4,t,K,n,arguments[2])},setFloat64:function(t,n){H(this,8,t,q,n,arguments[2])}});y(S,g),y(_,b),c(_[m],u.VIEW,!0),n[g]=S,n[b]=_},function(t,n,r){var e=r(3),i=r(52),o=r(69),u=r(182),c=r(11).f;t.exports=function(t){var n=i.Symbol||(i.Symbol=o?{}:e.Symbol||{});"_"==t.charAt(0)||t in n||c(n,t,{value:u.f(t)})}},function(t,n,r){var e=r(114),i=r(7)("iterator"),o=r(80);t.exports=r(52).getIteratorMethod=function(t){if(void 0!=t)return t[i]||t["@@iterator"]||o[e(t)]}},function(t,n,r){"use strict";var e=r(78),i=r(170),o=r(80),u=r(30);t.exports=r(140)(Array,"Array",function(t,n){this._t=u(t),this._i=0,this._k=n},function(){var t=this._t,n=this._k,r=this._i++;return!t||r>=t.length?(this._t=void 0,i(1)):"keys"==n?i(0,r):"values"==n?i(0,t[r]):i(0,[r,t[r]])},"values"),o.Arguments=o.Array,e("keys"),e("values"),e("entries")},function(t,n){function r(t,n){t.classList?t.classList.add(n):t.className+=" "+n}t.exports=r},function(t,n){function r(t,n){if(t.classList)t.classList.remove(n);else{var r=new RegExp("(^|\\b)"+n.split(" ").join("|")+"(\\b|$)","gi");t.className=t.className.replace(r," ")}}t.exports=r},function(t,n){function r(){throw new Error("setTimeout has not been defined")}function e(){throw new Error("clearTimeout has not been defined")}function i(t){if(s===setTimeout)return setTimeout(t,0);if((s===r||!s)&&setTimeout)return s=setTimeout,setTimeout(t,0);try{return s(t,0)}catch(n){try{return s.call(null,t,0)}catch(n){return s.call(this,t,0)}}}function o(t){if(l===clearTimeout)return clearTimeout(t);if((l===e||!l)&&clearTimeout)return l=clearTimeout,clearTimeout(t);try{return l(t)}catch(n){try{return l.call(null,t)}catch(n){return l.call(this,t)}}}function u(){d&&v&&(d=!1,v.length?p=v.concat(p):y=-1,p.length&&c())}function c(){if(!d){var t=i(u);d=!0;for(var n=p.length;n;){for(v=p,p=[];++y<n;)v&&v[y].run();y=-1,n=p.length}v=null,d=!1,o(t)}}function f(t,n){this.fun=t,this.array=n}function a(){}var s,l,h=t.exports={};
602!function(){try{s="function"==typeof setTimeout?setTimeout:r}catch(t){s=r}try{l="function"==typeof clearTimeout?clearTimeout:e}catch(t){l=e}}();var v,p=[],d=!1,y=-1;h.nextTick=function(t){var n=new Array(arguments.length-1);if(arguments.length>1)for(var r=1;r<arguments.length;r++)n[r-1]=arguments[r];p.push(new f(t,n)),1!==p.length||d||i(c)},f.prototype.run=function(){this.fun.apply(null,this.array)},h.title="browser",h.browser=!0,h.env={},h.argv=[],h.version="",h.versions={},h.on=a,h.addListener=a,h.once=a,h.off=a,h.removeListener=a,h.removeAllListeners=a,h.emit=a,h.prependListener=a,h.prependOnceListener=a,h.listeners=function(t){return[]},h.binding=function(t){throw new Error("process.binding is not supported")},h.cwd=function(){return"/"},h.chdir=function(t){throw new Error("process.chdir is not supported")},h.umask=function(){return 0}},function(t,n,r){var e=r(45);t.exports=function(t,n){if("number"!=typeof t&&"Number"!=e(t))throw TypeError(n);return+t}},function(t,n,r){"use strict";var e=r(17),i=r(75),o=r(16);t.exports=[].copyWithin||function(t,n){var r=e(this),u=o(r.length),c=i(t,u),f=i(n,u),a=arguments.length>2?arguments[2]:void 0,s=Math.min((void 0===a?u:i(a,u))-f,u-c),l=1;for(f<c&&c<f+s&&(l=-1,f+=s-1,c+=s-1);s-- >0;)f in r?r[c]=r[f]:delete r[c],c+=l,f+=l;return r}},function(t,n,r){var e=r(79);t.exports=function(t,n){var r=[];return e(t,!1,r.push,r,n),r}},function(t,n,r){var e=r(26),i=r(17),o=r(115),u=r(16);t.exports=function(t,n,r,c,f){e(n);var a=i(t),s=o(a),l=u(a.length),h=f?l-1:0,v=f?-1:1;if(r<2)for(;;){if(h in s){c=s[h],h+=v;break}if(h+=v,f?h<0:l<=h)throw TypeError("Reduce of empty array with no initial value")}for(;f?h>=0:l>h;h+=v)h in s&&(c=n(c,s[h],h,a));return c}},function(t,n,r){"use strict";var e=r(26),i=r(6),o=r(121),u=[].slice,c={},f=function(t,n,r){if(!(n in c)){for(var e=[],i=0;i<n;i++)e[i]="a["+i+"]";c[n]=Function("F,a","return new F("+e.join(",")+")")}return c[n](t,r)};t.exports=Function.bind||function(t){var n=e(this),r=u.call(arguments,1),c=function(){var e=r.concat(u.call(arguments));return this instanceof c?f(n,e.length,e):o(n,e,t)};return i(n.prototype)&&(c.prototype=n.prototype),c}},function(t,n,r){"use strict";var e=r(11).f,i=r(70),o=r(73),u=r(53),c=r(68),f=r(46),a=r(79),s=r(140),l=r(170),h=r(74),v=r(10),p=r(65).fastKey,d=v?"_s":"size",y=function(t,n){var r,e=p(n);if("F"!==e)return t._i[e];for(r=t._f;r;r=r.n)if(r.k==n)return r};t.exports={getConstructor:function(t,n,r,s){var l=t(function(t,e){c(t,l,n,"_i"),t._i=i(null),t._f=void 0,t._l=void 0,t[d]=0,void 0!=e&&a(e,r,t[s],t)});return o(l.prototype,{clear:function(){for(var t=this,n=t._i,r=t._f;r;r=r.n)r.r=!0,r.p&&(r.p=r.p.n=void 0),delete n[r.i];t._f=t._l=void 0,t[d]=0},delete:function(t){var n=this,r=y(n,t);if(r){var e=r.n,i=r.p;delete n._i[r.i],r.r=!0,i&&(i.n=e),e&&(e.p=i),n._f==r&&(n._f=e),n._l==r&&(n._l=i),n[d]--}return!!r},forEach:function(t){c(this,l,"forEach");for(var n,r=u(t,arguments.length>1?arguments[1]:void 0,3);n=n?n.n:this._f;)for(r(n.v,n.k,this);n&&n.r;)n=n.p},has:function(t){return!!y(this,t)}}),v&&e(l.prototype,"size",{get:function(){return f(this[d])}}),l},def:function(t,n,r){var e,i,o=y(t,n);return o?o.v=r:(t._l=o={i:i=p(n,!0),k:n,v:r,p:e=t._l,n:void 0,r:!1},t._f||(t._f=o),e&&(e.n=o),t[d]++,"F"!==i&&(t._i[i]=o)),t},getEntry:y,setStrong:function(t,n,r){s(t,n,function(t,n){this._t=t,this._k=n,this._l=void 0},function(){for(var t=this,n=t._k,r=t._l;r&&r.r;)r=r.p;return t._t&&(t._l=r=r?r.n:t._t._f)?"keys"==n?l(0,r.k):"values"==n?l(0,r.v):l(0,[r.k,r.v]):(t._t=void 0,l(1))},r?"entries":"values",!r,!0),h(n)}}},function(t,n,r){var e=r(114),i=r(161);t.exports=function(t){return function(){if(e(this)!=t)throw TypeError(t+"#toJSON isn't generic");return i(this)}}},function(t,n,r){"use strict";var e=r(73),i=r(65).getWeak,o=r(2),u=r(6),c=r(68),f=r(79),a=r(48),s=r(24),l=a(5),h=a(6),v=0,p=function(t){return t._l||(t._l=new d)},d=function(){this.a=[]},y=function(t,n){return l(t.a,function(t){return t[0]===n})};d.prototype={get:function(t){var n=y(this,t);if(n)return n[1]},has:function(t){return!!y(this,t)},set:function(t,n){var r=y(this,t);r?r[1]=n:this.a.push([t,n])},delete:function(t){var n=h(this.a,function(n){return n[0]===t});return~n&&this.a.splice(n,1),!!~n}},t.exports={getConstructor:function(t,n,r,o){var a=t(function(t,e){c(t,a,n,"_i"),t._i=v++,t._l=void 0,void 0!=e&&f(e,r,t[o],t)}
vendor: 4,224 bytes, line 602
602);return e(a.prototype,{delete:function(t){if(!u(t))return!1;var n=i(t);return!0===n?p(this).delete(t):n&&s(n,this._i)&&delete n[this._i]},has:function(t){if(!u(t))return!1;var n=i(t);return!0===n?p(this).has(t):n&&s(n,this._i)}}),a},def:function(t,n,r){var e=i(o(n),!0);return!0===e?p(t).set(n,r):e[t._i]=r,t},ufstore:p}},function(t,n,r){t.exports=!r(10)&&!r(4)(function(){return 7!=Object.defineProperty(r(132)("div"),"a",{get:function(){return 7}}).a})},function(t,n,r){var e=r(6),i=Math.floor;t.exports=function(t){return!e(t)&&isFinite(t)&&i(t)===t}},function(t,n,r){var e=r(2);t.exports=function(t,n,r,i){try{return i?n(e(r)[0],r[1]):n(r)}catch(n){var o=t.return;throw void 0!==o&&e(o.call(t)),n}}},function(t,n){t.exports=function(t,n){return{value:n,done:!!t}}},function(t,n){t.exports=Math.log1p||function(t){return(t=+t)>-1e-8&&t<1e-8?t-t*t/2:Math.log(1+t)}},function(t,n,r){"use strict";var e=r(72),i=r(125),o=r(116),u=r(17),c=r(115),f=Object.assign;t.exports=!f||r(4)(function(){var t={},n={},r=Symbol(),e="abcdefghijklmnopqrst";return t[r]=7,e.split("").forEach(function(t){n[t]=t}),7!=f({},t)[r]||Object.keys(f({},n)).join("")!=e})?function(t,n){for(var r=u(t),f=arguments.length,a=1,s=i.f,l=o.f;f>a;)for(var h,v=c(arguments[a++]),p=s?e(v).concat(s(v)):e(v),d=p.length,y=0;d>y;)l.call(v,h=p[y++])&&(r[h]=v[h]);return r}:f},function(t,n,r){var e=r(11),i=r(2),o=r(72);t.exports=r(10)?Object.defineProperties:function(t,n){i(t);for(var r,u=o(n),c=u.length,f=0;c>f;)e.f(t,r=u[f++],n[r]);return t}},function(t,n,r){var e=r(30),i=r(71).f,o={}.toString,u="object"==typeof window&&window&&Object.getOwnPropertyNames?Object.getOwnPropertyNames(window):[],c=function(t){try{return i(t)}catch(t){return u.slice()}};t.exports.f=function(t){return u&&"[object Window]"==o.call(t)?c(t):i(e(t))}},function(t,n,r){var e=r(24),i=r(30),o=r(117)(!1),u=r(145)("IE_PROTO");t.exports=function(t,n){var r,c=i(t),f=0,a=[];for(r in c)r!=u&&e(c,r)&&a.push(r);for(;n.length>f;)e(c,r=n[f++])&&(~o(a,r)||a.push(r));return a}},function(t,n,r){var e=r(72),i=r(30),o=r(116).f;t.exports=function(t){return function(n){for(var r,u=i(n),c=e(u),f=c.length,a=0,s=[];f>a;)o.call(u,r=c[a++])&&s.push(t?[r,u[r]]:u[r]);return s}}},function(t,n,r){var e=r(71),i=r(125),o=r(2),u=r(3).Reflect;t.exports=u&&u.ownKeys||function(t){var n=e.f(o(t)),r=i.f;return r?n.concat(r(t)):n}},function(t,n,r){var e=r(3).parseFloat,i=r(82).trim;t.exports=1/e(r(150)+"-0")!=-1/0?function(t){var n=i(String(t),3),r=e(n);return 0===r&&"-"==n.charAt(0)?-0:r}:e},function(t,n,r){var e=r(3).parseInt,i=r(82).trim,o=r(150),u=/^[\-+]?0[xX]/;t.exports=8!==e(o+"08")||22!==e(o+"0x16")?function(t,n){var r=i(String(t),3);return e(r,n>>>0||(u.test(r)?16:10))}:e},function(t,n){t.exports=Object.is||function(t,n){return t===n?0!==t||1/t==1/n:t!=t&&n!=n}},function(t,n,r){var e=r(16),i=r(149),o=r(46);t.exports=function(t,n,r,u){var c=String(o(t)),f=c.length,a=void 0===r?" ":String(r),s=e(n);if(s<=f||""==a)return c;var l=s-f,h=i.call(a,Math.ceil(l/a.length));return h.length>l&&(h=h.slice(0,l)),u?h+c:c+h}},function(t,n,r){n.f=r(7)},function(t,n,r){"use strict";var e=r(164);t.exports=r(118)("Map",function(t){return function(){return t(this,arguments.length>0?arguments[0]:void 0)}},{get:function(t){var n=e.getEntry(this,t);return n&&n.v},set:function(t,n){return e.def(this,0===t?0:t,n)}},e,!0)},function(t,n,r){r(10)&&"g"!=/./g.flags&&r(11).f(RegExp.prototype,"flags",{configurable:!0,get:r(120)})},function(t,n,r){"use strict";var e=r(164);t.exports=r(118)("Set",function(t){return function(){return t(this,arguments.length>0?arguments[0]:void 0)}},{add:function(t){return e.def(this,t=0===t?0:t,t)}},e)},function(t,n,r){"use strict";var e,i=r(48)(0),o=r(28),u=r(65),c=r(172),f=r(166),a=r(6),s=u.getWeak,l=Object.isExtensible,h=f.ufstore,v={},p=function(t){return function(){return t(this,arguments.length>0?arguments[0]:void 0)}},d={get:function(t){if(a(t)){var n=s(t);return!0===n?h(this).get(t):n?n[this._i]:void 0}},set:function(t,n){return f.def(this,t,n)}},y=t.exports=r(118)("WeakMap",p,d,f,!0,!0);7!=(new y).set((Object.freeze||Object)(v),7).get(v)&&(e=f.getConstructor(p),c(e.prototype,d),u.NEED=!0,i(["delete","has","get","set"],function(t){var n=y.prototype,r=n[t];
602o(n,t,function(n,i){if(a(n)&&!l(n)){this._f||(this._f=new e);var o=this._f[t](n,i);return"set"==t?this:o}return r.call(this,n,i)})}))},,,,function(t,n){"use strict";function r(){var t=document.querySelector("#page-nav");if(t&&!document.querySelector("#page-nav .extend.prev")&&(t.innerHTML='<a class="extend prev disabled" rel="prev">&laquo; Prev</a>'+t.innerHTML),t&&!document.querySelector("#page-nav .extend.next")&&(t.innerHTML=t.innerHTML+'<a class="extend next disabled" rel="next">Next &raquo;</a>'),yiliaConfig&&yiliaConfig.open_in_new){document.querySelectorAll(".article-entry a:not(.article-more-a)").forEach(function(t){var n=t.getAttribute("target");n&&""!==n||t.setAttribute("target","_blank")})}if(yiliaConfig&&yiliaConfig.toc_hide_index){document.querySelectorAll(".toc-number").forEach(function(t){t.style.display="none"})}var n=document.querySelector("#js-aboutme");n&&0!==n.length&&(n.innerHTML=n.innerText)}t.exports={init:r}},function(t,n,r){"use strict";function e(t){return t&&t.__esModule?t:{default:t}}function i(t,n){var r=/\/|index.html/g;return t.replace(r,"")===n.replace(r,"")}function o(){for(var t=document.querySelectorAll(".js-header-menu li a"),n=window.location.pathname,r=0,e=t.length;r<e;r++){var o=t[r];i(n,o.getAttribute("href"))&&(0,h.default)(o,"active")}}function u(t){for(var n=t.offsetLeft,r=t.offsetParent;null!==r;)n+=r.offsetLeft,r=r.offsetParent;return n}function c(t){for(var n=t.offsetTop,r=t.offsetParent;null!==r;)n+=r.offsetTop,r=r.offsetParent;return n}function f(t,n,r,e,i){var o=u(t),f=c(t)-n;if(f-r<=i){var a=t.$newDom;a||(a=t.cloneNode(!0),(0,d.default)(t,a),t.$newDom=a,a.style.position="fixed",a.style.top=(r||f)+"px",a.style.left=o+"px",a.style.zIndex=e||2,a.style.width="100%",a.style.color="#fff"),a.style.visibility="visible",t.style.visibility="hidden"}else{t.style.visibility="visible";var s=t.$newDom;s&&(s.style.visibility="hidden")}}function a(){var t=document.querySelector(".js-overlay"),n=document.querySelector(".js-header-menu");f(t,document.body.scrollTop,-63,2,0),f(n,document.body.scrollTop,1,3,0)}function s(){document.querySelector("#container").addEventListener("scroll",function(t){a()}),window.addEventListener("scroll",function(t){a()}),a()}var l=r(156),h=e(l),v=r(157),p=(e(v),r(382)),d=e(p),y=r(128),g=e(y),b=r(190),m=e(b),x=r(129);(function(){g.default.versions.mobile&&window.screen.width<800&&(o(),s())})(),(0,x.addLoadEvent)(function(){m.default.init()}),t.exports={}},,,,function(t,n,r){(function(t){"use strict";function n(t,n,r){t[n]||Object[e](t,n,{writable:!0,configurable:!0,value:r})}if(r(381),r(391),r(198),t._babelPolyfill)throw new Error("only one instance of babel-polyfill is allowed");t._babelPolyfill=!0;var e="defineProperty";n(String.prototype,"padLeft","".padStart),n(String.prototype,"padRight","".padEnd),"pop,reverse,shift,keys,values,entries,indexOf,every,some,forEach,map,filter,find,findIndex,includes,join,slice,concat,push,splice,unshift,sort,lastIndexOf,reduce,reduceRight,copyWithin,fill".split(",").forEach(function(t){[][t]&&n(Array,t,Function.call.bind([][t]))})}).call(n,function(){return this}())},,,function(t,n,r){r(210),t.exports=r(52).RegExp.escape},,,,function(t,n,r){var e=r(6),i=r(138),o=r(7)("species");t.exports=function(t){var n;return i(t)&&(n=t.constructor,"function"!=typeof n||n!==Array&&!i(n.prototype)||(n=void 0),e(n)&&null===(n=n[o])&&(n=void 0)),void 0===n?Array:n}},function(t,n,r){var e=r(202);t.exports=function(t,n){return new(e(t))(n)}},function(t,n,r){"use strict";var e=r(2),i=r(50),o="number";t.exports=function(t){if("string"!==t&&t!==o&&"default"!==t)throw TypeError("Incorrect hint");return i(e(this),t!=o)}},function(t,n,r){var e=r(72),i=r(125),o=r(116);t.exports=function(t){var n=e(t),r=i.f;if(r)for(var u,c=r(t),f=o.f,a=0;c.length>a;)f.call(t,u=c[a++])&&n.push(u);return n}},function(t,n,r){var e=r(72),i=r(30);t.exports=function(t,n){for(var r,o=i(t),u=e(o),c=u.length,f=0;c>f;)if(o[r=u[f++]]===n)return r}},function(t,n,r){"use strict";var e=r(208),i=r(121),o=r(26);t.exports=function(){for(var t=o(this),n=arguments.length,r=Array(n),u=0,c=e._,f=!1;n>u;)(r[u]=arguments[u++])===c&&(f=!0);return function(){var e,o=this,u=arguments.length,a=0,s=0;if(!f&&!u)return i(t,r,o);if(e=r.slice(),f)for(;n>a;a++)e[a]===c&&(e[a]=arguments[s++]);for(;u>s;)e.push(arguments[s++]);return i(t,e,o)}}},function(t,n,r){t.exports=r(3)},function(t,n){t.exports=function(t,n){var r=n===Object(n)?function(t){return n[t]}:n;return function(n){return String(n).replace(t,r)}}},function(t,n,r){var e=r(1),i=r(209)(/[\\^$*+?.()|[\]{}]/g,"\\$&");e(e.S,"RegExp",{escape:function(t){return i(t)}})},function(t,n,r){var e=r(1);e(e.P,"Array",{copyWithin:r(160)}),r(78)("copyWithin")},function(t,n,r){"use strict";var e=r(1),i=r(48)(4);e(e.P+e.F*!r(47)([].every,!0),"Array",{every:function(t){return i(this,t,arguments[1])}})},function(t,n,r){var e=r(1);e(e.P,"Array",{fill:r(130)}),r(78)("fill")},function(t,n,r){"use strict";var e=r(1),i=r(48)(2);e(e.P+e.F*!r(47)([].filter,!0),"Array",{filter:function(t){return i(this,t,arguments[1])}})},function(t,n,r){"use strict";var e=r(1),i=r(48)(6),o="findIndex",u=!0;o in[]&&Array(1)[o](function(){u=!1}),e(e.P+e.F*u,"Array",{findIndex:function(t){return i(this,t,arguments.length>1?arguments[1]:void 0)}}),r(78)(o)},function(t,n,r){"use strict";var e=r(1),i=r(48)(5),o="find",u=!0;o in[]&&Array(1)[o](function(){u=!1}),e(e.P+e.F*u,"Array",{find:function(t){return i(this,t,arguments.length>1?arguments[1]:void 0)}}),r(78)(o)},function(t,n,r){"use strict";var e=r(1),i=r(48)(0),o=r(47)([].forEach,!0);
602e(e.P+e.F*!o,"Array",{forEach:function(t){return i(this,t,arguments[1])}})},function(t,n,r){"use strict";var e=r(53),i=r(1),o=r(17),u=r(169),c=r(137),f=r(16),a=r(131),s=r(154);i(i.S+i.F*!r(123)(function(t){Array.from(t)}),"Array",{from:function(t){var n,r,i,l,h=o(t),v="function"==typeof this?this:Array,p=arguments.length,d=p>1?arguments[1]:void 0,y=void 0!==d,g=0,b=s(h);if(y&&(d=e(d,p>2?arguments[2]:void 0,2)),void 0==b||v==Array&&c(b))for(n=f(h.length),r=new v(n);n>g;g++)a(r,g,y?d(h[g],g):h[g]);else for(l=b.call(h),r=new v;!(i=l.next()).done;g++)a(r,g,y?u(l,d,[i.value,g],!0):i.value);return r.length=g,r}})},function(t,n,r){"use strict";var e=r(1),i=r(117)(!1),o=[].indexOf,u=!!o&&1/[1].indexOf(1,-0)<0;e(e.P+e.F*(u||!r(47)(o)),"Array",{indexOf:function(t){return u?o.apply(this,arguments)||0:i(this,t,arguments[1])}})},function(t,n,r){var e=r(1);e(e.S,"Array",{isArray:r(138)})},function(t,n,r){"use strict";var e=r(1),i=r(30),o=[].join;e(e.P+e.F*(r(115)!=Object||!r(47)(o)),"Array",{join:function(t){return o.call(i(this),void 0===t?",":t)}})},function(t,n,r){"use strict";var e=r(1),i=r(30),o=r(67),u=r(16),c=[].lastIndexOf,f=!!c&&1/[1].lastIndexOf(1,-0)<0;e(e.P+e.F*(f||!r(47)(c)),"Array",{lastIndexOf:function(t){if(f)return c.apply(this,arguments)||0;var n=i(this),r=u(n.length),e=r-1;for(arguments.length>1&&(e=Math.min(e,o(arguments[1]))),e<0&&(e=r+e);e>=0;e--)if(e in n&&n[e]===t)return e||0;return-1}})},function(t,n,r){"use strict";var e=r(1),i=r(48)(1);e(e.P+e.F*!r(47)([].map,!0),"Array",{map:function(t){return i(this,t,arguments[1])}})},function(t,n,r){"use strict";var e=r(1),i=r(131);e(e.S+e.F*r(4)(function(){function t(){}return!(Array.of.call(t)instanceof t)}),"Array",{of:function(){for(var t=0,n=arguments.length,r=new("function"==typeof this?this:Array)(n);n>t;)i(r,t,arguments[t++]);return r.length=n,r}})},function(t,n,r){"use strict";var e=r(1),i=r(162);e(e.P+e.F*!r(47)([].reduceRight,!0),"Array",{reduceRight:function(t){return i(this,t,arguments.length,arguments[1],!0)}})},function(t,n,r){"use strict";var e=r(1),i=r(162);e(e.P+e.F*!r(47)([].reduce,!0),"Array",{reduce:function(t){return i(this,t,arguments.length,arguments[1],!1)}})},function(t,n,r){"use strict";var e=r(1),i=r(135),o=r(45),u=r(75),c=r(16),f=[].slice;e(e.P+e.F*r(4)(function(){i&&f.call(i)}),"Array",{slice:function(t,n){var r=c(this.length),e=o(this);if(n=void 0===n?r:n,"Array"==e)return f.call(this,t,n);for(var i=u(t,r),a=u(n,r),s=c(a-i),l=Array(s),h=0;h<s;h++)l[h]="String"==e?this.charAt(i+h):this[i+h];return l}})},function(t,n,r){"use strict";var e=r(1),i=r(48)(3);e(e.P+e.F*!r(47)([].some,!0),"Array",{some:function(t){return i(this,t,arguments[1])}})},function(t,n,r){"use strict";var e=r(1),i=r(26),o=r(17),u=r(4),c=[].sort,f=[1,2,3];e(e.P+e.F*(u(function(){f.sort(void 0)})||!u(function(){f.sort(null)})||!r(47)(c)),"Array",{sort:function(t){return void 0===t?c.call(o(this)):c.call(o(this),i(t))}})},function(t,n,r){r(74)("Array")},function(t,n,r){var e=r(1);e(e.S,"Date",{now:function(){return(new Date).getTime()}})},function(t,n,r){"use strict";var e=r(1),i=r(4),o=Date.prototype.getTime,u=function(t){return t>9?t:"0"+t};e(e.P+e.F*(i(function(){return"0385-07-25T07:06:39.999Z"!=new Date(-5e13-1).toISOString()})||!i(function(){new Date(NaN).toISOString()})),"Date",{toISOString:function(){
603if(!isFinite(o.call(this)))throw RangeError("Invalid time value");var t=this,n=t.getUTCFullYear(),r=t.getUTCMilliseconds(),e=n<0?"-":n>9999?"+":"";return e+("00000"+Math.abs(n)).slice(e?-6:-4)+"-"+u(t.getUTCMonth()+1)+"-"+u(t.getUTCDate())+"T"+u(t.getUTCHours())+":"+u(t.getUTCMinutes())+":"+u(t.getUTCSeconds())+"."+(r>99?r:"0"+u(r))+"Z"}})},function(t,n,r){"use strict";var e=r(1),i=r(17),o=r(50);e(e.P+e.F*r(4)(function(){return null!==new Date(NaN).toJSON()||1!==Date.prototype.toJSON.call({toISOString:function(){return 1}})}),"Date",{toJSON:function(t){var n=i(this),r=o(n);return"number"!=typeof r||isFinite(r)?n.toISOString():null}})},function(t,n,r){var e=r(7)("toPrimitive"),i=Date.prototype;e in i||r(27)(i,e,r(204))},function(t,n,r){var e=Date.prototype,i="Invalid Date",o="toString",u=e[o],c=e.getTime;new Date(NaN)+""!=i&&r(28)(e,o,function(){var t=c.call(this);return t===t?u.call(this):i})},function(t,n,r){var e=r(1);e(e.P,"Function",{bind:r(163)})},function(t,n,r){"use strict";var e=r(6),i=r(32),o=r(7)("hasInstance"),u=Function.prototype;o in u||r(11).f(u,o,{value:function(t){if("function"!=typeof this||!e(t))return!1;if(!e(this.prototype))return t instanceof this;for(;t=i(t);)if(this.prototype===t)return!0;return!1}})},function(t,n,r){var e=r(11).f,i=r(66),o=r(24),u=Function.prototype,c="name",f=Object.isExtensible||function(){return!0};c in u||r(10)&&e(u,c,{configurable:!0,get:function(){try{var t=this,n=(""+t).match(/^\s*function ([^ (]*)/)[1];return o(t,c)||!f(t)||e(t,c,i(5,n)),n}catch(t){return""}}})},function(t,n,r){var e=r(1),i=r(171),o=Math.sqrt,u=Math.acosh;e(e.S+e.F*!(u&&710==Math.floor(u(Number.MAX_VALUE))&&u(1/0)==1/0),"Math",{acosh:function(t){return(t=+t)<1?NaN:t>94906265.62425156?Math.log(t)+Math.LN2:i(t-1+o(t-1)*o(t+1))}})},function(t,n,r){function e(t){return isFinite(t=+t)&&0!=t?t<0?-e(-t):Math.log(t+Math.sqrt(t*t+1)):t}var i=r(1),o=Math.asinh;i(i.S+i.F*!(o&&1/o(0)>0),"Math",{asinh:e})},function(t,n,r){var e=r(1),i=Math.atanh;e(e.S+e.F*!(i&&1/i(-0)<0),"Math",{atanh:function(t){return 0==(t=+t)?t:Math.log((1+t)/(1-t))/2}})},function(t,n,r){var e=r(1),i=r(142);e(e.S,"Math",{cbrt:function(t){return i(t=+t)*Math.pow(Math.abs(t),1/3)}})},function(t,n,r){var e=r(1);e(e.S,"Math",{clz32:function(t){return(t>>>=0)?31-Math.floor(Math.log(t+.5)*Math.LOG2E):32}})},function(t,n,r){var e=r(1),i=Math.exp;e(e.S,"Math",{cosh:function(t){return(i(t=+t)+i(-t))/2}})},function(t,n,r){var e=r(1),i=r(141);e(e.S+e.F*(i!=Math.expm1),"Math",{expm1:i})},function(t,n,r){var e=r(1),i=r(142),o=Math.pow,u=o(2,-52),c=o(2,-23),f=o(2,127)*(2-c),a=o(2,-126),s=function(t){return t+1/u-1/u};e(e.S,"Math",{fround:function(t){var n,r,e=Math.abs(t),o=i(t);return e<a?o*s(e/a/c)*a*c:(n=(1+c/u)*e,r=n-(n-e),r>f||r!=r?o*(1/0):o*r)}})},function(t,n,r){var e=r(1),i=Math.abs;e(e.S,"Math",{hypot:function(t,n){for(var r,e,o=0,u=0,c=arguments.length,f=0;u<c;)r=i(arguments[u++]),f<r?(e=f/r,o=o*e*e+1,f=r):r>0?(e=r/f,o+=e*e):o+=r;return f===1/0?1/0:f*Math.sqrt(o)}})},function(t,n,r){var e=r(1),i=Math.imul;e(e.S+e.F*r(4)(function(){return-5!=i(4294967295,5)||2!=i.length}),"Math",{imul:function(t,n){var r=65535,e=+t,i=+n,o=r&e,u=r&i;return 0|o*u+((r&e>>>16)*u+o*(r&i>>>16)<<16>>>0)}})},function(t,n,r){var e=r(1);e(e.S,"Math",{log10:function(t){return Math.log(t)/Math.LN10}})},function(t,n,r){var e=r(1);e(e.S,"Math",{log1p:r(171)})},function(t,n,r){var e=r(1);e(e.S,"Math",{log2:function(t){return Math.log(t)/Math.LN2}})},function(t,n,r){var e=r(1);e(e.S,"Math",{sign:r(142)})},function(t,n,r){var e=r(1),i=r(141),o=Math.exp;e(e.S+e.F*r(4)(function(){return-2e-17!=!Math.sinh(-2e-17)}),"Math",{sinh:function(t){return Math.abs(t=+t)<1?(i(t)-i(-t))/2:(o(t-1)-o(-t-1))*(Math.E/2)}})},function(t,n,r){var e=r(1),i=r(141),o=Math.exp;e(e.S,"Math",{tanh:function(t){var n=i(t=+t),r=i(-t);return n==1/0?1:r==1/0?-1:(n-r)/(o(t)+o(-t))}})},function(t,n,r){var e=r(1);e(e.S,"Math",{trunc:function(t){return(t>0?Math.floor:Math.ceil)(t)}})},function(t,n,r){"use strict";var e=r(3),i=r(24),o=r(45),u=r(136),c=r(50),f=r(4),a=r(71).f,s=r(31).f,l=r(11).f,h=r(82).trim,v="Number",p=e[v],d=p,y=p.prototype,g=o(r(70)(y))==v,b="trim"in String.prototype,m=function(t){var n=c(t,!1);if("string"==typeof n&&n.length>2){n=b?n.trim():h(n,3);var r,e,i,o=n.charCodeAt(0);if(43===o||45===o){if(88===(r=n.charCodeAt(2))||120===r)return NaN}else if(48===o){switch(n.charCodeAt(1)){case 66:case 98:e=2,i=49;break;case 79:case 111:e=8,i=55;break;default:return+n}for(var u,f=n.slice(2),a=0,s=f.length;a<s;a++)if((u=f.charCodeAt(a))<48||u>i)return NaN;return parseInt(f,e)}}return+n};if(!p(" 0o1")||!p("0b1")||p("+0x1")){p=function(t){var n=arguments.length<1?0:t,r=this;return r instanceof p&&(g?f(function(){y.valueOf.call(r)}):o(r)!=v)?u(new d(m(n)),r,p):m(n)};for(var x,w=r(10)?a(d):"MAX_VALUE,MIN_VALUE,NaN,NEGATIVE_INFINITY,POSITIVE_INFINITY,EPSILON,isFinite,isInteger,isNaN,isSafeInteger,MAX_SAFE_INTEGER,MIN_SAFE_INTEGER,parseFloat,parseInt,isInteger".split(","),S=0;w.length>S;S++)i(d,x=w[S])&&!i(p,x)&&l(p,x,s(d,x));p.prototype=y,y.constructor=p,r(28)(e,v,p)}},function(t,n,r){var e=r(1);e(e.S,"Number",{EPSILON:Math.pow(2,-52)})},function(t,n,r){var e=r(1),i=r(3).isFinite;e(e.S,"Number",{isFinite:function(t){return"number"==typeof t&&i(t)}})},function(t,n,r){var e=r(1);e(e.S,"Number",{isInteger:r(168)})},function(t,n,r){var e=r(1);e(e.S,"Number",{isNaN:function(t){return t!=t}})},function(t,n,r){var e=r(1),i=r(168),o=Math.abs;e(e.S,"Number",{isSafeInteger:function(t){return i(t)&&o(t)<=9007199254740991}})},function(t,n,r){var e=r(1);e(e.S,"Number",{MAX_SAFE_INTEGER:9007199254740991})},function(t,n,r){var e=r(1);e(e.S,"Number",{MIN_SAFE_INTEGER:-9007199254740991})},function(t,n,r){var e=r(1),i=r(178);e(e.S+e.F*(Number.parseFloat!=i),"Number",{parseFloat:i})},function(t,n,r){var e=r(1),i=r(179);
603e(e.S+e.F*(Number.parseInt!=i),"Number",{parseInt:i})},function(t,n,r){"use strict";var e=r(1),i=r(67),o=r(159),u=r(149),c=1..toFixed,f=Math.floor,a=[0,0,0,0,0,0],s="Number.toFixed: incorrect invocation!",l="0",h=function(t,n){for(var r=-1,e=n;++r<6;)e+=t*a[r],a[r]=e%1e7,e=f(e/1e7)},v=function(t){for(var n=6,r=0;--n>=0;)r+=a[n],a[n]=f(r/t),r=r%t*1e7},p=function(){for(var t=6,n="";--t>=0;)if(""!==n||0===t||0!==a[t]){var r=String(a[t]);n=""===n?r:n+u.call(l,7-r.length)+r}return n},d=function(t,n,r){return 0===n?r:n%2==1?d(t,n-1,r*t):d(t*t,n/2,r)},y=function(t){for(var n=0,r=t;r>=4096;)n+=12,r/=4096;for(;r>=2;)n+=1,r/=2;return n};e(e.P+e.F*(!!c&&("0.000"!==8e-5.toFixed(3)||"1"!==.9.toFixed(0)||"1.25"!==1.255.toFixed(2)||"1000000000000000128"!==(0xde0b6b3a7640080).toFixed(0))||!r(4)(function(){c.call({})})),"Number",{toFixed:function(t){var n,r,e,c,f=o(this,s),a=i(t),g="",b=l;if(a<0||a>20)throw RangeError(s);if(f!=f)return"NaN";if(f<=-1e21||f>=1e21)return String(f);if(f<0&&(g="-",f=-f),f>1e-21)if(n=y(f*d(2,69,1))-69,r=n<0?f*d(2,-n,1):f/d(2,n,1),r*=4503599627370496,(n=52-n)>0){for(h(0,r),e=a;e>=7;)h(1e7,0),e-=7;for(h(d(10,e,1),0),e=n-1;e>=23;)v(1<<23),e-=23;v(1<<e),h(1,1),v(2),b=p()}else h(0,r),h(1<<-n,0),b=p()+u.call(l,a);return a>0?(c=b.length,b=g+(c<=a?"0."+u.call(l,a-c)+b:b.slice(0,c-a)+"."+b.slice(c-a))):b=g+b,b}})},function(t,n,r){"use strict";var e=r(1),i=r(4),o=r(159),u=1..toPrecision;e(e.P+e.F*(i(function(){return"1"!==u.call(1,void 0)})||!i(function(){u.call({})})),"Number",{toPrecision:function(t){var n=o(this,"Number#toPrecision: incorrect invocation!");return void 0===t?u.call(n):u.call(n,t)}})},function(t,n,r){var e=r(1);e(e.S+e.F,"Object",{assign:r(172)})},function(t,n,r){var e=r(1);e(e.S,"Object",{create:r(70)})},function(t,n,r){var e=r(1);e(e.S+e.F*!r(10),"Object",{defineProperties:r(173)})},function(t,n,r){var e=r(1);e(e.S+e.F*!r(10),"Object",{defineProperty:r(11).f})},function(t,n,r){var e=r(6),i=r(65).onFreeze;r(49)("freeze",function(t){return function(n){return t&&e(n)?t(i(n)):n}})},function(t,n,r){var e=r(30),i=r(31).f;r(49)("getOwnPropertyDescriptor",function(){return function(t,n){return i(e(t),n)}})},function(t,n,r){r(49)("getOwnPropertyNames",function(){return r(174).f})},function(t,n,r){var e=r(17),i=r(32);r(49)("getPrototypeOf",function(){return function(t){return i(e(t))}})},function(t,n,r){var e=r(6);r(49)("isExtensible",function(t){return function(n){return!!e(n)&&(!t||t(n))}})},function(t,n,r){var e=r(6);r(49)("isFrozen",function(t){return function(n){return!e(n)||!!t&&t(n)}})},function(t,n,r){var e=r(6);r(49)("isSealed",function(t){return function(n){return!e(n)||!!t&&t(n)}})},function(t,n,r){var e=r(1);e(e.S,"Object",{is:r(180)})},function(t,n,r){var e=r(17),i=r(72);r(49)("keys",function(){return function(t){return i(e(t))}})},function(t,n,r){var e=r(6),i=r(65).onFreeze;r(49)("preventExtensions",function(t){return function(n){return t&&e(n)?t(i(n)):n}})},function(t,n,r){var e=r(6),i=r(65).onFreeze;r(49)("seal",function(t){return function(n){return t&&e(n)?t(i(n)):n}})},function(t,n,r){var e=r(1);e(e.S,"Object",{setPrototypeOf:r(144).set})},function(t,n,r){"use strict";var e=r(114),i={};i[r(7)("toStringTag")]="z",i+""!="[object z]"&&r(28)(Object.prototype,"toString",function(){return"[object "+e(this)+"]"},!0)},function(t,n,r){var e=r(1),i=r(178);e(e.G+e.F*(parseFloat!=i),{parseFloat:i})},function(t,n,r){var e=r(1),i=r(179);e(e.G+e.F*(parseInt!=i),{parseInt:i})},function(t,n,r){"use strict";var e,i,o,u=r(69),c=r(3),f=r(53),a=r(114),s=r(1),l=r(6),h=r(26),v=r(68),p=r(79),d=r(146),y=r(151).set,g=r(143)(),b="Promise",m=c.TypeError,x=c.process,w=c[b],x=c.process,S="process"==a(x),_=function(){},O=!!function(){try{var t=w.resolve(1),n=(t.constructor={})[r(7)("species")]=function(t){t(_,_)};return(S||"function"==typeof PromiseRejectionEvent)&&t.then(_)instanceof n}catch(t){}}(),E=function(t,n){return t===n||t===w&&n===o},P=function(t){var n;return!(!l(t)||"function"!=typeof(n=t.then))&&n},j=function(t){return E(w,t)?new F(t):new i(t)},F=i=function(t){var n,r;this.promise=new t(function(t,e){if(void 0!==n||void 0!==r)throw m("Bad Promise constructor");n=t,r=e}),this.resolve=h(n),this.reject=h(r)},M=function(t){try{t()}catch(t){return{error:t}}},A=function(t,n){if(!t._n){t._n=!0;var r=t._c;g(function(){for(var e=t._v,i=1==t._s,o=0;r.length>
603o;)!function(n){var r,o,u=i?n.ok:n.fail,c=n.resolve,f=n.reject,a=n.domain;try{u?(i||(2==t._h&&I(t),t._h=1),!0===u?r=e:(a&&a.enter(),r=u(e),a&&a.exit()),r===n.promise?f(m("Promise-chain cycle")):(o=P(r))?o.call(r,c,f):c(r)):f(e)}catch(t){f(t)}}(r[o++]);t._c=[],t._n=!1,n&&!t._h&&N(t)})}},N=function(t){y.call(c,function(){var n,r,e,i=t._v;if(T(t)&&(n=M(function(){S?x.emit("unhandledRejection",i,t):(r=c.onunhandledrejection)?r({promise:t,reason:i}):(e=c.console)&&e.error&&e.error("Unhandled promise rejection",i)}),t._h=S||T(t)?2:1),t._a=void 0,n)throw n.error})},T=function(t){if(1==t._h)return!1;for(var n,r=t._a||t._c,e=0;r.length>e;)if(n=r[e++],n.fail||!T(n.promise))return!1;return!0},I=function(t){y.call(c,function(){var n;S?x.emit("rejectionHandled",t):(n=c.onrejectionhandled)&&n({promise:t,reason:t._v})})},k=function(t){var n=this;n._d||(n._d=!0,n=n._w||n,n._v=t,n._s=2,n._a||(n._a=n._c.slice()),A(n,!0))},L=function(t){var n,r=this;if(!r._d){r._d=!0,r=r._w||r;try{if(r===t)throw m("Promise can't be resolved itself");(n=P(t))?g(function(){var e={_w:r,_d:!1};try{n.call(t,f(L,e,1),f(k,e,1))}catch(t){k.call(e,t)}}):(r._v=t,r._s=1,A(r,!1))}catch(t){k.call({_w:r,_d:!1},t)}}};O||(w=function(t){v(this,w,b,"_h"),h(t),e.call(this);try{t(f(L,this,1),f(k,this,1))}catch(t){k.call(this,t)}},e=function(t){this._c=[],this._a=void 0,this._s=0,this._d=!1,this._v=void 0,this._h=0,this._n=!1},e.prototype=r(73)(w.prototype,{then:function(t,n){var r=j(d(this,w));return r.ok="function"!=typeof t||t,r.fail="function"==typeof n&&n,r.domain=S?x.domain:void 0,this._c.push(r),this._a&&this._a.push(r),this._s&&A(this,!1),r.promise},catch:function(t){return this.then(void 0,t)}}),F=function(){var t=new e;this.promise=t,this.resolve=f(L,t,1),this.reject=f(k,t,1)}),s(s.G+s.W+s.F*!O,{Promise:w}),r(81)(w,b),r(74)(b),o=r(52)[b],s(s.S+s.F*!O,b,{reject:function(t){var n=j(this);return(0,n.reject)(t),n.promise}}),s(s.S+s.F*(u||!O),b,{resolve:function(t){if(t instanceof w&&E(t.constructor,this))return t;var n=j(this);return(0,n.resolve)(t),n.promise}}),s(s.S+s.F*!(O&&r(123)(function(t){w.all(t).catch(_)})),b,{all:function(t){var n=this,r=j(n),e=r.resolve,i=r.reject,o=M(function(){var r=[],o=0,u=1;p(t,!1,function(t){var c=o++,f=!1;r.push(void 0),u++,n.resolve(t).then(function(t){f||(f=!0,r[c]=t,--u||e(r))},i)}),--u||e(r)});return o&&i(o.error),r.promise},race:function(t){var n=this,r=j(n),e=r.reject,i=M(function(){p(t,!1,function(t){n.resolve(t).then(r.resolve,e)})});return i&&e(i.error),r.promise}})},function(t,n,r){var e=r(1),i=r(26),o=r(2),u=(r(3).Reflect||{}).apply,c=Function.apply;e(e.S+e.F*!r(4)(function(){u(function(){})}),"Reflect",{apply:function(t,n,r){var e=i(t),f=o(r);return u?u(e,n,f):c.call(e,n,f)}})},function(t,n,r){var e=r(1),i=r(70),o=r(26),u=r(2),c=r(6),f=r(4),a=r(163),s=(r(3).Reflect||{}).construct,l=f(function(){function t(){}return!(s(function(){},[],t)instanceof t)}),h=!f(function(){s(function(){})});e(e.S+e.F*(l||h),"Reflect",{construct:function(t,n){o(t),u(n);var r=arguments.length<3?t:o(arguments[2]);if(h&&!l)return s(t,n,r);if(t==r){switch(n.length){case 0:return new t;case 1:return new t(n[0]);case 2:return new t(n[0],n[1]);case 3:return new t(n[0],n[1],n[2]);case 4:return new t(n[0],n[1],n[2],n[3])}var e=[null];return e.push.apply(e,n),new(a.apply(t,e))}var f=r.prototype,v=i(c(f)?f:Object.prototype),p=Function.apply.call(t,v,n);return c(p)?p:v}})},function(t,n,r){var e=r(11),i=r(1),o=r(2),u=r(50);i(i.S+i.F*r(4)(function(){Reflect.defineProperty(e.f({},1,{value:1}),1,{value:2})}),"Reflect",{defineProperty:function(t,n,r){o(t),n=u(n,!0),o(r);try{return e.f(t,n,r),!0}catch(t){return!1}}})},function(t,n,r){var e=r(1),i=r(31).f,o=r(2);e(e.S,"Reflect",{deleteProperty:function(t,n){var r=i(o(t),n);return!(r&&!r.configurable)&&delete t[n]}})},function(t,n,r){"use strict";var e=r(1),i=r(2),o=function(t){this._t=i(t),this._i=0;var n,r=this._k=[];for(n in t)r.push(n)};r(139)(o,"Object",function(){var t,n=this,r=n._k;do{if(n._i>=r.length)return{value:void 0,done:!0}}while(!((t=r[n._i++])in n._t));return{value:t,done:!1}}),e(e.S,"Reflect",{enumerate:function(t){return new o(t)}})},function(t,n,r){var e=r(31),i=r(1),o=r(2);i(i.S,"Reflect",{getOwnPropertyDescriptor:function(t,n){return e.f(o(t),n)}})},function(t,n,r){var e=r(1),i=r(32),o=r(2);e(e.S,"Reflect",{getPrototypeOf:function(t){return i(o(t))}})},function(t,n,r){function e(t,n){var r,c,s=arguments.length<3?t:arguments[2];return a(t)===s?t[n]:(r=i.f(t,n))?u(r,"value")?r.value:void 0!==r.get?r.get.call(s):void 0:f(c=o(t))?e(c,n,s):void 0}var i=r(31),o=r(32),u=r(24),c=r(1),f=r(6),a=r(2);c(c.S,"Reflect",{get:e})},function(t,n,r){var e=r(1);e(e.S,"Reflect",{has:function(t,n){return n in t}})},function(t,n,r){var e=r(1),i=r(2),o=Object.isExtensible;e(e.S,"Reflect",{isExtensible:function(t){return i(t),!o||o(t)}})},function(t,n,r){var e=r(1);e(e.S,"Reflect",{ownKeys:r(177)})},function(t,n,r){var e=r(1),i=r(2),o=Object.preventExtensions;e(e.S,"Reflect",{preventExtensions:function(t){i(t);try{return o&&o(t),!0}catch(t){return!1}}})},function(t,n,r){var e=r(1),i=r(144);i&&e(e.S,"Reflect",{setPrototypeOf:function(t,n){i.check(t,n);try{return i.set(t,n),!0}catch(t){return!1}}})},function(t,n,r){function e(t,n,r){var f,h,v=arguments.length<4?t:
603arguments[3],p=o.f(s(t),n);if(!p){if(l(h=u(t)))return e(h,n,r,v);p=a(0)}return c(p,"value")?!(!1===p.writable||!l(v)||(f=o.f(v,n)||a(0),f.value=r,i.f(v,n,f),0)):void 0!==p.set&&(p.set.call(v,r),!0)}var i=r(11),o=r(31),u=r(32),c=r(24),f=r(1),a=r(66),s=r(2),l=r(6);f(f.S,"Reflect",{set:e})},function(t,n,r){var e=r(3),i=r(136),o=r(11).f,u=r(71).f,c=r(122),f=r(120),a=e.RegExp,s=a,l=a.prototype,h=/a/g,v=/a/g,p=new a(h)!==h;if(r(10)&&(!p||r(4)(function(){return v[r(7)("match")]=!1,a(h)!=h||a(v)==v||"/a/i"!=a(h,"i")}))){a=function(t,n){var r=this instanceof a,e=c(t),o=void 0===n;return!r&&e&&t.constructor===a&&o?t:i(p?new s(e&&!o?t.source:t,n):s((e=t instanceof a)?t.source:t,e&&o?f.call(t):n),r?this:l,a)};for(var d=u(s),y=0;d.length>y;)!function(t){t in a||o(a,t,{configurable:!0,get:function(){return s[t]},set:function(n){s[t]=n}})}(d[y++]);l.constructor=a,a.prototype=l,r(28)(e,"RegExp",a)}r(74)("RegExp")},function(t,n,r){r(119)("match",1,function(t,n,r){return[function(r){"use strict";var e=t(this),i=void 0==r?void 0:r[n];return void 0!==i?i.call(r,e):new RegExp(r)[n](String(e))},r]})},function(t,n,r){r(119)("replace",2,function(t,n,r){return[function(e,i){"use strict";var o=t(this),u=void 0==e?void 0:e[n];return void 0!==u?u.call(e,o,i):r.call(String(o),e,i)},r]})},function(t,n,r){r(119)("search",1,function(t,n,r){return[function(r){"use strict";var e=t(this),i=void 0==r?void 0:r[n];return void 0!==i?i.call(r,e):new RegExp(r)[n](String(e))},r]})},function(t,n,r){r(119)("split",2,function(t,n,e){"use strict";var i=r(122),o=e,u=[].push,c="split",f="length",a="lastIndex";if("c"=="abbc"[c](/(b)*/)[1]||4!="test"[c](/(?:)/,-1)[f]||2!="ab"[c](/(?:ab)*/)[f]||4!="."[c](/(.?)(.?)/)[f]||"."[c](/()()/)[f]>1||""[c](/.?/)[f]){var s=void 0===/()??/.exec("")[1];e=function(t,n){var r=String(this);if(void 0===t&&0===n)return[];if(!i(t))return o.call(r,t,n);var e,c,l,h,v,p=[],d=(t.ignoreCase?"i":"")+(t.multiline?"m":"")+(t.unicode?"u":"")+(t.sticky?"y":""),y=0,g=void 0===n?4294967295:n>>>0,b=new RegExp(t.source,d+"g");for(s||(e=new RegExp("^"+b.source+"$(?!\\s)",d));(c=b.exec(r))&&!((l=c.index+c[0][f])>y&&(p.push(r.slice(y,c.index)),!s&&c[f]>1&&c[0].replace(e,function(){for(v=1;v<arguments[f]-2;v++)void 0===arguments[v]&&(c[v]=void 0)}),c[f]>1&&c.index<r[f]&&u.apply(p,c.slice(1)),h=c[0][f],y=l,p[f]>=g));)b[a]===c.index&&b[a]++;return y===r[f]?!h&&b.test("")||p.push(""):p.push(r.slice(y)),p[f]>g?p.slice(0,g):p}}else"0"[c](void 0,0)[f]&&(e=function(t,n){return void 0===t&&0===n?[]:o.call(this,t,n)});return[function(r,i){var o=t(this),u=void 0==r?void 0:r[n];return void 0!==u?u.call(r,o,i):e.call(String(o),r,i)},e]})},function(t,n,r){"use strict";r(184);var e=r(2),i=r(120),o=r(10),u="toString",c=/./[u],f=function(t){r(28)(RegExp.prototype,u,t,!0)};r(4)(function(){return"/a/b"!=c.call({source:"a",flags:"b"})})?f(function(){var t=e(this);return"/".concat(t.source,"/","flags"in t?t.flags:!o&&t instanceof RegExp?i.call(t):void 0)}):c.name!=u&&f(function(){return c.call(this)})},function(t,n,r){"use strict";r(29)("anchor",function(t){return function(n){return t(this,"a","name",n)}})},function(t,n,r){"use strict";r(29)("big",function(t){return function(){return t(this,"big","","")}})},function(t,n,r){"use strict";r(29)("blink",function(t){return function(){return t(this,"blink","","")}})},function(t,n,r){"use strict";r(29)("bold",function(t){return function(){return t(this,"b","","")}})},function(t,n,r){"use strict";var e=r(1),i=r(147)(!1);e(e.P,"String",{codePointAt:function(t){return i(this,t)}})},function(t,n,r){"use strict";var e=r(1),i=r(16),o=r(148),u="endsWith",c=""[u];e(e.P+e.F*r(134)(u),"String",{endsWith:function(t){var n=o(this,t,u),r=arguments.length>1?arguments[1]:void 0,e=i(n.length),f=void 0===r?e:Math.min(i(r),e),a=String(t);return c?c.call(n,a,f):n.slice(f-a.length,f)===a}})},function(t,n,r){"use strict";r(29)("fixed",function(t){return function(){return t(this,"tt","","")}})},function(t,n,r){"use strict";r(29)("fontcolor",function(t){return function(n){return t(this,"font","color",n)}})},function(t,n,r){"use strict";r(29)("fontsize",function(t){return function(n){return t(this,"font","size",n)}})},function(t,n,r){var e=r(1),i=r(75),o=String.fromCharCode,u=String.fromCodePoint;e(e.S+e.F*(!!u&&1!=u.length),"String",{fromCodePoint:function(t){for(var n,r=[],e=arguments.length,u=0;e>u;){if(n=+arguments[u++],i(n,1114111)!==n)throw RangeError(n+" is not a valid code point");r.push(n<65536?o(n):o(55296+((n-=65536)>>10),n%1024+56320))}return r.join("")}})},function(t,n,r){"use strict";var e=r(1),i=r(148),o="includes";e(e.P+e.F*r(134)(o),"String",{includes:function(t){return!!~i(this,t,o).indexOf(t,arguments.length>1?arguments[1]:void 0)}})},function(t,n,r){"use strict";r(29)("italics",function(t){return function(){return t(this,"i","","")}})},function(t,n,r){"use strict";var e=r(147)(!0);
vendor: 14,837 bytes, lines 603-604
603r(140)(String,"String",function(t){this._t=String(t),this._i=0},function(){var t,n=this._t,r=this._i;return r>=n.length?{value:void 0,done:!0}:(t=e(n,r),this._i+=t.length,{value:t,done:!1})})},function(t,n,r){"use strict";r(29)("link",function(t){return function(n){return t(this,"a","href",n)}})},function(t,n,r){var e=r(1),i=r(30),o=r(16);e(e.S,"String",{raw:function(t){for(var n=i(t.raw),r=o(n.length),e=arguments.length,u=[],c=0;r>c;)u.push(String(n[c++])),c<e&&u.push(String(arguments[c]));return u.join("")}})},function(t,n,r){var e=r(1);e(e.P,"String",{repeat:r(149)})},function(t,n,r){"use strict";r(29)("small",function(t){return function(){return t(this,"small","","")}})},function(t,n,r){"use strict";var e=r(1),i=r(16),o=r(148),u="startsWith",c=""[u];e(e.P+e.F*r(134)(u),"String",{startsWith:function(t){var n=o(this,t,u),r=i(Math.min(arguments.length>1?arguments[1]:void 0,n.length)),e=String(t);return c?c.call(n,e,r):n.slice(r,r+e.length)===e}})},function(t,n,r){"use strict";r(29)("strike",function(t){return function(){return t(this,"strike","","")}})},function(t,n,r){"use strict";r(29)("sub",function(t){return function(){return t(this,"sub","","")}})},function(t,n,r){"use strict";r(29)("sup",function(t){return function(){return t(this,"sup","","")}})},function(t,n,r){"use strict";r(82)("trim",function(t){return function(){return t(this,3)}})},function(t,n,r){"use strict";var e=r(3),i=r(24),o=r(10),u=r(1),c=r(28),f=r(65).KEY,a=r(4),s=r(126),l=r(81),h=r(76),v=r(7),p=r(182),d=r(153),y=r(206),g=r(205),b=r(138),m=r(2),x=r(30),w=r(50),S=r(66),_=r(70),O=r(174),E=r(31),P=r(11),j=r(72),F=E.f,M=P.f,A=O.f,N=e.Symbol,T=e.JSON,I=T&&T.stringify,k="prototype",L=v("_hidden"),R=v("toPrimitive"),C={}.propertyIsEnumerable,D=s("symbol-registry"),U=s("symbols"),W=s("op-symbols"),G=Object[k],B="function"==typeof N,V=e.QObject,z=!V||!V[k]||!V[k].findChild,q=o&&a(function(){return 7!=_(M({},"a",{get:function(){return M(this,"a",{value:7}).a}})).a})?function(t,n,r){var e=F(G,n);e&&delete G[n],M(t,n,r),e&&t!==G&&M(G,n,e)}:M,K=function(t){var n=U[t]=_(N[k]);return n._k=t,n},J=B&&"symbol"==typeof N.iterator?function(t){return"symbol"==typeof t}:function(t){return t instanceof N},Y=function(t,n,r){return t===G&&Y(W,n,r),m(t),n=w(n,!0),m(r),i(U,n)?(r.enumerable?(i(t,L)&&t[L][n]&&(t[L][n]=!1),r=_(r,{enumerable:S(0,!1)})):(i(t,L)||M(t,L,S(1,{})),t[L][n]=!0),q(t,n,r)):M(t,n,r)},H=function(t,n){m(t);for(var r,e=g(n=x(n)),i=0,o=e.length;o>i;)Y(t,r=e[i++],n[r]);return t},$=function(t,n){return void 0===n?_(t):H(_(t),n)},X=function(t){var n=C.call(this,t=w(t,!0));return!(this===G&&i(U,t)&&!i(W,t))&&(!(n||!i(this,t)||!i(U,t)||i(this,L)&&this[L][t])||n)},Q=function(t,n){if(t=x(t),n=w(n,!0),t!==G||!i(U,n)||i(W,n)){var r=F(t,n);return!r||!i(U,n)||i(t,L)&&t[L][n]||(r.enumerable=!0),r}},Z=function(t){for(var n,r=A(x(t)),e=[],o=0;r.length>o;)i(U,n=r[o++])||n==L||n==f||e.push(n);return e},tt=function(t){for(var n,r=t===G,e=A(r?W:x(t)),o=[],u=0;e.length>u;)!i(U,n=e[u++])||r&&!i(G,n)||o.push(U[n]);return o};B||(N=function(){if(this instanceof N)throw TypeError("Symbol is not a constructor!");var t=h(arguments.length>0?arguments[0]:void 0),n=function(r){this===G&&n.call(W,r),i(this,L)&&i(this[L],t)&&(this[L][t]=!1),q(this,t,S(1,r))};return o&&z&&q(G,t,{configurable:!0,set:n}),K(t)},c(N[k],"toString",function(){return this._k}),E.f=Q,P.f=Y,r(71).f=O.f=Z,r(116).f=X,r(125).f=tt,o&&!r(69)&&c(G,"propertyIsEnumerable",X,!0),p.f=function(t){return K(v(t))}),u(u.G+u.W+u.F*!B,{Symbol:N});for(var nt="hasInstance,isConcatSpreadable,iterator,match,replace,search,species,split,toPrimitive,toStringTag,unscopables".split(","),rt=0;nt.length>rt;)v(nt[rt++]);for(var nt=j(v.store),rt=0;nt.length>rt;)d(nt[rt++]);u(u.S+u.F*!B,"Symbol",{for:function(t){return i(D,t+="")?D[t]:D[t]=N(t)},keyFor:function(t){if(J(t))return y(D,t);throw TypeError(t+" is not a symbol!")},useSetter:function(){z=!0},useSimple:function(){z=!1}}),u(u.S+u.F*!B,"Object",{create:$,defineProperty:Y,defineProperties:H,getOwnPropertyDescriptor:Q,getOwnPropertyNames:Z,getOwnPropertySymbols:tt}),T&&u(u.S+u.F*(!B||a(function(){var t=N();return"[null]"!=I([t])||"{}"!=I({a:t})||"{}"!=I(Object(t))})),"JSON",{stringify:function(t){if(void 0!==t&&!J(t)){for(var n,r,e=[t],i=1;arguments.length>i;)e.push(arguments[i++]);return n=e[1],"function"==typeof n&&(r=n),!r&&b(n)||(n=function(t,n){if(r&&(n=r.call(this,t,n)),!J(n))return n}),e[1]=n,I.apply(T,e)}}}),N[k][R]||r(27)(N[k],R,N[k].valueOf),l(N,"Symbol"),l(Math,"Math",!0),l(e.JSON,"JSON",!0)},function(t,n,r){"use strict";var e=r(1),i=r(127),o=r(152),u=r(2),c=r(75),f=r(16),a=r(6),s=r(3).ArrayBuffer,l=r(146),h=o.ArrayBuffer,v=o.DataView,p=i.ABV&&s.isView,d=h.prototype.slice,y=i.VIEW,g="ArrayBuffer";e(e.G+e.W+e.F*(s!==h),{ArrayBuffer:h}),e(e.S+e.F*!i.CONSTR,g,{isView:function(t){return p&&p(t)||a(t)&&y in t}}),e(e.P+e.U+e.F*r(4)(function(){return!new h(2).slice(1,void 0).byteLength}),g,{slice:function(t,n){if(void 0!==d&&void 0===n)return d.call(u(this),t);for(var r=u(this).byteLength,e=c(t,r),i=c(void 0===n?r:n,r),o=new(l(this,h))(f(i-e)),a=new v(this),s=new v(o),p=0;e<i;)s.setUint8(p++,a.getUint8(e++));return o}}),r(74)(g)},function(t,n,r){var e=r(1);e(e.G+e.W+e.F*!r(127).ABV,{DataView:r(152).DataView})},function(t,n,r){r(55)("Float32",4,function(t){return function(n,r,e){return t(this,n,r,e)}})},function(t,n,r){r(55)("Float64",8,function(t){return function(n,r,e){return t(this,n,r,e)}})},function(t,n,r){r(55)("Int16",2,function(t){return function(n,r,e){return t(this,n,r,e)}})},function(t,n,r){r(55)("Int32",4,function(t){return function(n,r,e){return t(this,n,r,e)}})},function(t,n,r){r(55)("Int8",1,function(t){return function(n,r,e){return t(this,n,r,e)}})},function(t,n,r){r(55)("Uint16",2,function(t){return function(n,r,e){return t(this,n,r,e)}})},function(t,n,r){r(55)("Uint32",4,function(t){return function(n,r,e){return t(this,n,r,e)}})},function(t,n,r){r(55)("Uint8",1,function(t){return function(n,r,e){return t(this,n,r,e)}})},function(t,n,r){r(55)("Uint8",1,function(t){return function(n,r,e){return t(this,n,r,e)}},!0)},function(t,n,r){"use strict";var e=r(166);r(118)("WeakSet",function(t){return function(){return t(this,arguments.length>0?arguments[0]:void 0)}},{add:function(t){return e.def(this,t,!0)}},e,!1,!0)},function(t,n,r){"use strict";var e=r(1),i=r(117)(!0);e(e.P,"Array",{includes:function(t){return i(this,t,arguments.length>1?arguments[1]:void 0)}}),r(78)("includes")},function(t,n,r){var e=r(1),i=r(143)(),o=r(3).process,u="process"==r(45)(o);e(e.G,{asap:function(t){var n=u&&o.domain;i(n?n.bind(t):t)}})},function(t,n,r){var e=r(1),i=r(45);e(e.S,"Error",{isError:function(t){return"Error"===i(t)}})},function(t,n,r){var e=r(1);e(e.P+e.R,"Map",{toJSON:r(165)("Map")})},function(t,n,r){var e=r(1);e(e.S,"Math",{iaddh:function(t,n,r,e){var i=t>>>0,o=n>>>0,u=r>>>0;return o+(e>>>0)+((i&u|(i|u)&~(i+u>>>0))>>>31)|0}})},function(t,n,r){var e=r(1);e(e.S,"Math",{imulh:function(t,n){var r=65535,e=+t,i=+n,o=e&r,u=i&r,c=e>>16,f=i>>16,a=(c*u>>>0)+(o*u>>>16);return c*f+(a>>16)+((o*f>>>0)+(a&r)>>16)}})},function(t,n,r){var e=r(1);e(e.S,"Math",{isubh:function(t,n,r,e){var i=t>>>0,o=n>>>0,u=r>>>0;return o-(e>>>0)-((~i&u|~(i^u)&i-u>>>0)>>>31)|0}})},function(t,n,r){var e=r(1);e(e.S,"Math",{umulh:function(t,n){var r=65535,e=+t,i=+n,o=e&r,u=i&r,c=e>>>16,f=i>>>16,a=(c*u>>>0)+(o*u>>>16);return c*f+(a>>>16)+((o*f>>>0)+(a&r)>>>16)}})},function(t,n,r){"use strict";var e=r(1),i=r(17),o=r(26),u=r(11);r(10)&&e(e.P+r(124),"Object",{__defineGetter__:function(t,n){u.f(i(this),t,{get:o(n),enumerable:!0,configurable:!0})}})},function(t,n,r){"use strict";var e=r(1),i=r(17),o=r(26),u=r(11);r(10)&&e(e.P+r(124),"Object",{__defineSetter__:function(t,n){u.f(i(this),t,{set:o(n),enumerable:!0,configurable:!0})}})},function(t,n,r){var e=r(1),i=r(176)(!0);e(e.S,"Object",{entries:function(t){return i(t)}})},function(t,n,r){var e=r(1),i=r(177),o=r(30),u=r(31),c=r(131);e(e.S,"Object",{getOwnPropertyDescriptors:function(t){for(var n,r=o(t),e=u.f,f=i(r),a={},s=0;f.length>s;)c(a,n=f[s++],e(r,n));return a}})},function(t,n,r){"use strict";var e=r(1),i=r(17),o=r(50),u=r(32),c=r(31).f;r(10)&&e(e.P+r(124),"Object",{__lookupGetter__:function(t){var n,r=i(this),e=o(t,!0);do{if(n=c(r,e))return n.get}while(r=u(r))}})},function(t,n,r){"use strict";var e=r(1),i=r(17),o=r(50),u=r(32),c=r(31).f;r(10)&&e(e.P+r(124),"Object",{__lookupSetter__:function(t){var n,r=i(this),e=o(t,!0);do{if(n=c(r,e))return n.set}while(r=u(r))}})},function(t,n,r){var e=r(1),i=r(176)(!1);e(e.S,"Object",{values:function(t){return i(t)}})},function(t,n,r){"use strict";var e=r(1),i=r(3),o=r(52),u=r(143)(),c=r(7)("observable"),f=r(26),a=r(2),s=r(68),l=r(73),h=r(27),v=r(79),p=v.RETURN,d=function(t){return null==t?void 0:f(t)},y=function(t){var n=t._c;n&&(t._c=void 0,n())},g=function(t){return void 0===t._o},b=function(t){g(t)||(t._o=void 0,y(t))},m=function(t,n){a(t),this._c=void 0,this._o=t,t=new x(this);try{var r=n(t),e=r;null!=r&&("function"==typeof r.unsubscribe?r=function(){e.unsubscribe()}:f(r),this._c=r)}catch(n){return void t.error(n)}g(this)&&y(this)};m.prototype=l({},{unsubscribe:function(){b(this)}});var x=function(t){this._s=t};x.prototype=l({},{next:function(t){var n=this._s;if(!g(n)){var r=n._o;try{var e=d(r.next);if(e)return e.call(r,t)}catch(t){try{b(n)}finally{throw t}}}},error:function(t){var n=this._s;if(g(n))throw t;var r=n._o;n._o=void 0;try{var e=d(r.error);if(!e)throw t;t=e.call(r,t)}catch(t){try{y(n)}finally{throw t}}return y(n),t},complete:function(t){var n=this._s;if(!g(n)){var r=n._o;n._o=void 0;try{var e=d(r.complete);t=e?e.call(r,t):void 0}catch(t){try{y(n)}finally{throw t}}return y(n),t}}});var w=function(t){s(this,w,"Observable","_f")._f=f(t)};l(w.prototype,{subscribe:function(t){return new m(t,this._f)},forEach:function(t){var n=this;return new(o.Promise||i.Promise)(function(r,e){f(t);var i=n.subscribe({next:function(n){try{return t(n)}catch(t){e(t),i.unsubscribe()}},error:e,complete:r})})}}),l(w,{from:function(t){var n="function"==typeof this?this:w,r=d(a(t)[c]);if(r){var e=a(r.call(t));return e.constructor===n?e:new n(function(t){return e.subscribe(t)})}return new n(function(n){var r=!1;return u(function(){if(!r){try{if(v(t,!1,function(t){if(n.next(t),r)return p})===p)return}catch(t){if(r)throw t;return void n.error(t)}n.complete()}}),function(){r=!0}})},of:function(){for(var t=0,n=arguments.length,r=Array(n);t<n;)r[t]=arguments[t++];return new("function"==typeof this?this:w)(function(t){var n=!1;return u(function(){if(!n){for(var e=0;e<r.length;++e)if(t.next(r[e]),n)return;t.complete()}}),function(){n=!0}})}}),h(w.prototype,c,function(){return this}),e(e.G,{Observable:w}),r(74)("Observable")},function(t,n,r){var e=r(54),i=r(2),o=e.key,u=e.set;e.exp({defineMetadata:function(t,n,r,e){u(t,n,i(r),o(e))}})},function(t,n,r){var e=r(54),i=r(2),o=e.key,u=e.map,c=e.store;e.exp({deleteMetadata:function(t,n){var r=arguments.length<3?void 0:o(arguments[2]),e=u(i(n),r,!1);if(void 0===e||!e.delete(t))return!1;if(e.size)return!0;var f=c.get(n);return f.delete(r),!!f.size||c.delete(n)}})},function(t,n,r){var e=r(185),i=r(161),o=r(54),u=r(2),c=r(32),f=o.keys,a=o.key,s=function(t,n){var r=f(t,n),o=c(t);if(null===o)return r;var u=s(o,n);return u.length?r.length?i(new e(r.concat(u))):u:r};o.exp({getMetadataKeys:function(t){return s(u(t),arguments.length<2?void 0:a(arguments[1]))}})},function(t,n,r){var e=r(54),i=r(2),o=r(32),u=e.has,c=e.get,f=e.key,a=function(t,n,r){if(u(t,n,r))return c(t,n,r);var e=o(n);return null!==e?a(t,e,r):void 0};e.exp({getMetadata:function(t,n){return a(t,i(n),arguments.length<3?void 0:f(arguments[2]))}})},function(t,n,r){var e=r(54),i=r(2),o=e.keys,u=e.key;e.exp({getOwnMetadataKeys:function(t){
604return o(i(t),arguments.length<2?void 0:u(arguments[1]))}})},function(t,n,r){var e=r(54),i=r(2),o=e.get,u=e.key;e.exp({getOwnMetadata:function(t,n){return o(t,i(n),arguments.length<3?void 0:u(arguments[2]))}})},function(t,n,r){var e=r(54),i=r(2),o=r(32),u=e.has,c=e.key,f=function(t,n,r){if(u(t,n,r))return!0;var e=o(n);return null!==e&&f(t,e,r)};e.exp({hasMetadata:function(t,n){return f(t,i(n),arguments.length<3?void 0:c(arguments[2]))}})},function(t,n,r){var e=r(54),i=r(2),o=e.has,u=e.key;e.exp({hasOwnMetadata:function(t,n){return o(t,i(n),arguments.length<3?void 0:u(arguments[2]))}})},function(t,n,r){var e=r(54),i=r(2),o=r(26),u=e.key,c=e.set;e.exp({metadata:function(t,n){return function(r,e){c(t,n,(void 0!==e?i:o)(r),u(e))}}})},function(t,n,r){var e=r(1);e(e.P+e.R,"Set",{toJSON:r(165)("Set")})},function(t,n,r){"use strict";var e=r(1),i=r(147)(!0);e(e.P,"String",{at:function(t){return i(this,t)}})},function(t,n,r){"use strict";var e=r(1),i=r(46),o=r(16),u=r(122),c=r(120),f=RegExp.prototype,a=function(t,n){this._r=t,this._s=n};r(139)(a,"RegExp String",function(){var t=this._r.exec(this._s);return{value:t,done:null===t}}),e(e.P,"String",{matchAll:function(t){if(i(this),!u(t))throw TypeError(t+" is not a regexp!");var n=String(this),r="flags"in f?String(t.flags):c.call(t),e=new RegExp(t.source,~r.indexOf("g")?r:"g"+r);return e.lastIndex=o(t.lastIndex),new a(e,n)}})},function(t,n,r){"use strict";var e=r(1),i=r(181);e(e.P,"String",{padEnd:function(t){return i(this,t,arguments.length>1?arguments[1]:void 0,!1)}})},function(t,n,r){"use strict";var e=r(1),i=r(181);e(e.P,"String",{padStart:function(t){return i(this,t,arguments.length>1?arguments[1]:void 0,!0)}})},function(t,n,r){"use strict";r(82)("trimLeft",function(t){return function(){return t(this,1)}},"trimStart")},function(t,n,r){"use strict";r(82)("trimRight",function(t){return function(){return t(this,2)}},"trimEnd")},function(t,n,r){r(153)("asyncIterator")},function(t,n,r){r(153)("observable")},function(t,n,r){var e=r(1);e(e.S,"System",{global:r(3)})},function(t,n,r){for(var e=r(155),i=r(28),o=r(3),u=r(27),c=r(80),f=r(7),a=f("iterator"),s=f("toStringTag"),l=c.Array,h=["NodeList","DOMTokenList","MediaList","StyleSheetList","CSSRuleList"],v=0;v<5;v++){var p,d=h[v],y=o[d],g=y&&y.prototype;if(g){g[a]||u(g,a,l),g[s]||u(g,s,d),c[d]=l;for(p in e)g[p]||i(g,p,e[p],!0)}}},function(t,n,r){var e=r(1),i=r(151);e(e.G+e.B,{setImmediate:i.set,clearImmediate:i.clear})},function(t,n,r){var e=r(3),i=r(1),o=r(121),u=r(207),c=e.navigator,f=!!c&&/MSIE .\./.test(c.userAgent),a=function(t){return f?function(n,r){return t(o(u,[].slice.call(arguments,2),"function"==typeof n?n:Function(n)),r)}:t};i(i.G+i.B+i.F*f,{setTimeout:a(e.setTimeout),setInterval:a(e.setInterval)})},function(t,n,r){r(330),r(269),r(271),r(270),r(273),r(275),r(280),r(274),r(272),r(282),r(281),r(277),r(278),r(276),r(268),r(279),r(283),r(284),r(236),r(238),r(237),r(286),r(285),r(256),r(266),r(267),r(257),r(258),r(259),r(260),r(261),r(262),r(263),r(264),r(265),r(239),r(240),r(241),r(242),r(243),r(244),r(245),r(246),r(247),r(248),r(249),r(250),r(251),r(252),r(253),r(254),r(255
vendor: 7,469 bytes, line 604
604),r(317),r(322),r(329),r(320),r(312),r(313),r(318),r(323),r(325),r(308),r(309),r(310),r(311),r(314),r(315),r(316),r(319),r(321),r(324),r(326),r(327),r(328),r(231),r(233),r(232),r(235),r(234),r(220),r(218),r(224),r(221),r(227),r(229),r(217),r(223),r(214),r(228),r(212),r(226),r(225),r(219),r(222),r(211),r(213),r(216),r(215),r(230),r(155),r(302),r(307),r(184),r(303),r(304),r(305),r(306),r(287),r(183),r(185),r(186),r(342),r(331),r(332),r(337),r(340),r(341),r(335),r(338),r(336),r(339),r(333),r(334),r(288),r(289),r(290),r(291),r(292),r(295),r(293),r(294),r(296),r(297),r(298),r(299),r(301),r(300),r(343),r(369),r(372),r(371),r(373),r(374),r(370),r(375),r(376),r(354),r(357),r(353),r(351),r(352),r(355),r(356),r(346),r(368),r(377),r(345),r(347),r(349),r(348),r(350),r(359),r(360),r(362),r(361),r(364),r(363),r(365),r(366),r(367),r(344),r(358),r(380),r(379),r(378),t.exports=r(52)},function(t,n){function r(t,n){if("string"==typeof n)return t.insertAdjacentHTML("afterend",n);var r=t.nextSibling;return r?t.parentNode.insertBefore(n,r):t.parentNode.appendChild(n)}t.exports=r},,,,,,,,,function(t,n,r){(function(n,r){!function(n){"use strict";function e(t,n,r,e){var i=n&&n.prototype instanceof o?n:o,u=Object.create(i.prototype),c=new p(e||[]);return u._invoke=s(t,r,c),u}function i(t,n,r){try{return{type:"normal",arg:t.call(n,r)}}catch(t){return{type:"throw",arg:t}}}function o(){}function u(){}function c(){}function f(t){["next","throw","return"].forEach(function(n){t[n]=function(t){return this._invoke(n,t)}})}function a(t){function n(r,e,o,u){var c=i(t[r],t,e);if("throw"!==c.type){var f=c.arg,a=f.value;return a&&"object"==typeof a&&m.call(a,"__await")?Promise.resolve(a.__await).then(function(t){n("next",t,o,u)},function(t){n("throw",t,o,u)}):Promise.resolve(a).then(function(t){f.value=t,o(f)},u)}u(c.arg)}function e(t,r){function e(){return new Promise(function(e,i){n(t,r,e,i)})}return o=o?o.then(e,e):e()}"object"==typeof r&&r.domain&&(n=r.domain.bind(n));var o;this._invoke=e}function s(t,n,r){var e=P;return function(o,u){if(e===F)throw new Error("Generator is already running");if(e===M){if("throw"===o)throw u;return y()}for(r.method=o,r.arg=u;;){var c=r.delegate;if(c){var f=l(c,r);if(f){if(f===A)continue;return f}}if("next"===r.method)r.sent=r._sent=r.arg;else if("throw"===r.method){if(e===P)throw e=M,r.arg;r.dispatchException(r.arg)}else"return"===r.method&&r.abrupt("return",r.arg);e=F;var a=i(t,n,r);if("normal"===a.type){if(e=r.done?M:j,a.arg===A)continue;return{value:a.arg,done:r.done}}"throw"===a.type&&(e=M,r.method="throw",r.arg=a.arg)}}}function l(t,n){var r=t.iterator[n.method];if(r===g){if(n.delegate=null,"throw"===n.method){if(t.iterator.return&&(n.method="return",n.arg=g,l(t,n),"throw"===n.method))return A;n.method="throw",n.arg=new TypeError("The iterator does not provide a 'throw' method")}return A}var e=i(r,t.iterator,n.arg);if("throw"===e.type)return n.method="throw",n.arg=e.arg,n.delegate=null,A;var o=e.arg;return o?o.done?(n[t.resultName]=o.value,n.next=t.nextLoc,"return"!==n.method&&(n.method="next",n.arg=g),n.delegate=null,A):o:(n.method="throw",n.arg=new TypeError("iterator result is not an object"),n.delegate=null,A)}function h(t){var n={tryLoc:t[0]};1 in t&&(n.catchLoc=t[1]),2 in t&&(n.finallyLoc=t[2],n.afterLoc=t[3]),this.tryEntries.push(n)}function v(t){var n=t.completion||{};n.type="normal",delete n.arg,t.completion=n}function p(t){this.tryEntries=[{tryLoc:"root"}],t.forEach(h,this),this.reset(!0)}function d(t){if(t){var n=t[w];if(n)return n.call(t);if("function"==typeof t.next)return t;if(!isNaN(t.length)){var r=-1,e=function n(){for(;++r<t.length;)if(m.call(t,r))return n.value=t[r],n.done=!1,n;return n.value=g,n.done=!0,n};return e.next=e}}return{next:y}}function y(){return{value:g,done:!0}}var g,b=Object.prototype,m=b.hasOwnProperty,x="function"==typeof Symbol?Symbol:{},w=x.iterator||"@@iterator",S=x.asyncIterator||"@@asyncIterator",_=x.toStringTag||"@@toStringTag",O="object"==typeof t,E=n.regeneratorRuntime;if(E)return void(O&&(t.exports=E));E=n.regeneratorRuntime=O?t.exports:{},E.wrap=e;var P="suspendedStart",j="suspendedYield",F="executing",M="completed",A={},N={};N[w]=function(){return this};var T=Object.getPrototypeOf,I=T&&T(T(d([])));I&&I!==b&&m.call(I,w)&&(N=I);var k=c.prototype=o.prototype=Object.create(N);u.prototype=k.constructor=c,c.constructor=u,c[_]=u.displayName="GeneratorFunction",E.isGeneratorFunction=function(t){var n="function"==typeof t&&t.constructor;return!!n&&(n===u||"GeneratorFunction"===(n.displayName||n.name))},E.mark=function(t){return Object.setPrototypeOf?Object.setPrototypeOf(t,c):(t.__proto__=c,_ in t||(t[_]="GeneratorFunction")),t.prototype=Object.create(k),t},E.awrap=function(t){return{__await:t}},f(a.prototype),a.prototype[S]=function(){return this},E.AsyncIterator=a,E.async=function(t,n,r,i){var o=new a(e(t,n,r,i));return E.isGeneratorFunction(n)?o:o.next().then(function(t){return t.done?t.value:o.next()})},f(k),k[_]="Generator",k.toString=function(){return"[object Generator]"},E.keys=function(t){var n=[];for(var r in t)n.push(r);return n.reverse(),function r(){for(;n.length;){var e=n.pop();if(e in t)return r.value=e,r.done=!1,r}return r.done=!0,r}},E.values=d,p.prototype={constructor:p,reset:function(t){if(this.prev=0,this.next=0,this.sent=this._sent=g,this.done=!1,this.delegate=null,this.method="next",this.arg=g,this.tryEntries.forEach(v),!t)for(var n in this)"t"===n.charAt(0)&&m.call(this,n)&&!isNaN(+n.slice(1))&&(this[n]=g)},stop:function(){this.done=!0;var t=this.tryEntries[0],n=t.completion;if("throw"===n.type)throw n.arg;return this.rval},dispatchException:function(t){function n(n,e){return o.type="throw",o.arg=t,r.next=n,e&&(r.method="next",r.arg=g),!!e}if(this.done)throw t;for(var r=this,e=this.tryEntries.length-1;e>=0;--e){var i=this.tryEntries[e],o=i.completion;if("root"===i.tryLoc)return n("end");if(i.tryLoc<=this.prev){var u=m.call(i,"catchLoc"),c=m.call(i,"finallyLoc");if(u&&c){if(this.prev<i.catchLoc)return n(i.catchLoc,!0);if(this.prev<i.finallyLoc)return n(i.finallyLoc)}else if(u){if(this.prev<i.catchLoc)return n(i.catchLoc,!0)}else{if(!c)throw new Error("try statement without catch or finally");if(this.prev<i.finallyLoc)return n(i.finallyLoc)}}}},abrupt:function(t,n){for(var r=this.tryEntries.length-1;r>=0;--r){var e=this.tryEntries[r];if(e.tryLoc<=this.prev&&m.call(e,"finallyLoc")&&this.prev<e.finallyLoc){var i=e;break}}i&&("break"===t||"continue"===t)&&i.tryLoc<=n&&n<=i.finallyLoc&&(i=null);var o=i?i.completion:{};return o.type=t,o.arg=n,i?(this.method="next",this.next=i.finallyLoc,A):this.complete(o)},complete:function(t,n){if("throw"===t.type)throw t.arg;return"break"===t.type||"continue"===t.type?this.next=t.arg:"return"===t.type?(this.rval=this.arg=t.arg,this.method="return",this.next="end"):"normal"===t.type&&n&&(this.next=n),A},finish:function(t){for(var n=this.tryEntries.length-1;n>=0;--n){var r=this.tryEntries[n];if(r.finallyLoc===t)return this.complete(r.completion,r.afterLoc),v(r),A}},catch:function(t){for(var n=this.tryEntries.length-1;n>=0;--n){var r=this.tryEntries[n];if(r.tryLoc===t){var e=r.completion;if("throw"===e.type){var i=e.arg;v(r)}return i}}throw new Error("illegal catch attempt")},delegateYield:function(t,n,r){return this.delegate={iterator:d(t),resultName:n,nextLoc:r},"next"===this.method&&(this.arg=g),A}}}("object"==typeof n?n:"object"==typeof window?window:"object"==typeof self?self:this)}).call(n,function(){return this}(),r(158))}])</script>
604<script src="/./main.0cf68a.js"></script>
604<script>!function(){!function(e){var t=document.createElement("script");document.getElementsByTagName("body")[0].appendChild(t),t.setAttribute("src",e)}("/slider.e37972.js")}()</script>
604
605
606
607<script type="text/x-mathjax-config">
608MathJax.Hub.Config({
609    tex2jax: {
610        inlineMath: [ ['$','$'], ["\\(","\\)"]  ],
611        processEscapes: true,
612        skipTags: ['script', 'noscript', 'style', 'textarea', 'pre', 'code']
613    }
614});
615
616MathJax.Hub.Queue(function() {
617    var all = MathJax.Hub.getAllJax(), i;
618    for(i=0; i < all.length; i += 1) {
619        all[i].SourceElement().parentNode.className += ' has-jax';                 
620    }       
621});
622</script>
622
623
624<script src="//cdn.mathjax.org/mathjax/latest/MathJax.js?config=TeX-AMS-MML_HTMLorMML">
625</script>
625
626
627
628    
629<div class="tools-col" q-class="show:isShow,hide:isShow|isFalse" q-on="click:stop(e)">
630  <div class="tools-nav header-menu">
631    
632    
633      
634      
635      
636    
637      
638      
639      
640    
641      
642      
643      
644    
645    
646
647    <ul style="width: 70%">
648    
649    
650      
651      <li style="width: 33.333333333333336%" q-on="click: openSlider(e, 'innerArchive')"><a href="javascript:void(0)" q-class="active:innerArchive">所有文章</a></li>
652      
653        
654      
655      <li style="width: 33.333333333333336%" q-on="click: openSlider(e, 'friends')"><a href="javascript:void(0)" q-class="active:friends">友情链接</a></li>
656      
657        
658      
659      <li style="width: 33.333333333333336%" q-on="click: openSlider(e, 'aboutme')"><a href="javascript:void(0)" q-class="active:aboutme">关于我</a></li>
660      
661        
662    </ul>
663  </div>
664  <div class="tools-wrap">
665    
666    	<section class="tools-section tools-section-all" q-show="innerArchive">
667        <div class="search-wrap">
668          <input class="search-ipt" q-model="search" type="text" placeholder="find something…">
669          <i class="icon-search icon" q-show="search|isEmptyStr"></i>
670          <i class="icon-close icon" q-show="search|isNotEmptyStr" q-on="click:clearChose(e)"></i>
671        </div>
672        <div class="widget tagcloud search-tag">
673          <p class="search-tag-wording">tag:</p>
674          <label class="search-switch">
675            <input type="checkbox" q-on="click:toggleTag(e)" q-attr="checked:showTags">
676          </label>
677          <ul class="article-tag-list" q-show="showTags">
678             
679              <li class="article-tag-list-item">
680                <a href="javascript:void(0)" class="js-tag color2">DeepFM</a>
681              </li>
682             
683              <li class="article-tag-list-item">
684                <a href="javascript:void(0)" class="js-tag color3">FM</a>
685              </li>
686             
687              <li class="article-tag-list-item">
688                <a href="javascript:void(0)" class="js-tag color4">CTR</a>
689              </li>
690             
691              <li class="article-tag-list-item">
692                <a href="javascript:void(0)" class="js-tag color2">Github</a>
693              </li>
694             
695              <li class="article-tag-list-item">
696                <a href="javascript:void(0)" class="js-tag color5">Microsoft</a>
697              </li>
698             
699              <li class="article-tag-list-item">
700                <a href="javascript:void(0)" class="js-tag color5">hexo</a>
701              </li>
702             
703              <li class="article-tag-list-item">
704                <a href="javascript:void(0)" class="js-tag color2">github</a>
705              </li>
706             
707              <li class="article-tag-list-item">
708                <a href="javascript:void(0)" class="js-tag color5">搭建博客</a>
709              </li>
710             
711              <li class="article-tag-list-item">
712                <a href="javascript:void(0)" class="js-tag color1">LatexTools</a>
713              </li>
714             
715              <li class="article-tag-list-item">
716                <a href="javascript:void(0)" class="js-tag color3">sublime</a>
717              </li>
718             
719              <li class="article-tag-list-item">
720                <a href="javascript:void(0)" class="js-tag color5">参考文献</a>
721              </li>
722             
723              <li class="article-tag-list-item">
724                <a href="javascript:void(0)" class="js-tag color4">ocr</a>
725              </li>
726             
727              <li class="article-tag-list-item">
728                <a href="javascript:void(0)" class="js-tag color5">CTPN</a>
729              </li>
730             
731              <li class="article-tag-list-item">
732                <a href="javascript:void(0)" class="js-tag color2">图像文本检测</a>
733              </li>
734             
735              <li class="article-tag-list-item">
736                <a href="javascript:void(0)" class="js-tag color3">haskell</a>
737              </li>
738             
739              <li class="article-tag-list-item">
740                <a href="javascript:void(0)" class="js-tag color2">python</a>
741              </li>
742             
743              <li class="article-tag-list-item">
744                <a href="javascript:void(0)" class="js-tag color4">PIL</a>
745              </li>
746             
747              <li class="article-tag-list-item">
748                <a href="javascript:void(0)" class="js-tag color5">图像处理</a>
749              </li>
750             
751              <li class="article-tag-list-item">
752                <a href="javascript:void(0)" class="js-tag color4">gif</a>
753              </li>
754             
755              <li class="article-tag-list-item">
756                <a href="javascript:void(0)" class="js-tag color1">shell</a>
757              </li>
758             
759              <li class="article-tag-list-item">
760                <a href="javascript:void(0)" class="js-tag color4">GAN</a>
761              </li>
762             
763              <li class="article-tag-list-item">
764                <a href="javascript:void(0)" class="js-tag color1">条件GAN</a>
765              </li>
766             
767              <li class="article-tag-list-item">
768                <a href="javascript:void(0)" class="js-tag color2">生成对抗网络</a>
769              </li>
770             
771              <li class="article-tag-list-item">
772                <a href="javascript:void(0)" class="js-tag color5">论文解读</a>
773              </li>
774             
775              <li class="article-tag-list-item">
776                <a href="javascript:void(0)" class="js-tag color5">sopt</a>
777              </li>
778             
779              <li class="article-tag-list-item">
780                <a href="javascript:void(0)" class="js-tag color4">最优化</a>
781              </li>
782             
783              <li class="article-tag-list-item">
784                <a href="javascript:void(0)" class="js-tag color3">optimization</a>
785              </li>
786             
787              <li class="article-tag-list-item">
788                <a href="javascript:void(0)" class="js-tag color1">DCGAN</a>
789              </li>
790             
791              <li class="article-tag-list-item">
792                <a href="javascript:void(0)" class="js-tag color5">代码解读</a>
793              </li>
794            
795            <div class="clearfix"></div>
796          </ul>
797        </div>
798        <ul class="search-ul">
799          <p q-show="jsonFail" style="padding: 20px; font-size: 12px;">
800            缺失模块。<br/>1、请确保node版本大于6.2<br/>2、在博客根目录(注意不是yilia根目录)执行以下命令:<br/> npm i hexo-generator-json-content --save<br/><br/>
801            3、在根目录_config.yml里添加配置:
802<pre style="font-size: 12px;" q-show="jsonFail">
803  jsonContent:
804    meta: false
805    pages: false
806    posts:
807      title: true
808      date: true
809      path: true
810      text: false
811      raw: false
812      content: false
813      slug: false
814      updated: false
815      comments: false
816      link: false
817      permalink: false
818      excerpt: false
819      categories: false
820      tags: true
821</pre>
822          </p>
823          <li class="search-li" q-repeat="items" q-show="isShow">
824            <a q-attr="href:path|urlformat" class="search-title"><i class="icon-quo-left icon"></i><span q-text="title"></span></a>
825            <p class="search-time">
826              <i class="icon-calendar icon"></i>
827              <span q-text="date|dateformat"></span>
828            </p>
829            <p class="search-tag">
830              <i class="icon-price-tags icon"></i>
831              <span q-repeat="tags" q-on="click:choseTag(e, name)" q-text="name|tagformat"></span>
832            </p>
833          </li>
834        </ul>
835    	</section>
836    
837
838    
839    	<section class="tools-section tools-section-friends" q-show="friends">
840  		
841        <ul class="search-ul">
842          
843            <li class="search-li">
844              <a href="http://www.github.com/Lyrichu" target="_blank" class="search-title"><i class="icon-quo-left icon"></i>github</a>
845            </li>
846          
847            <li class="search-li">
848              <a href="http://www.movieb2b.com" target="_blank" class="search-title"><i class="icon-quo-left icon"></i>个人网页</a>
849            </li>
850          
851            <li class="search-li">
852              <a href="http://www.cnblogs.com/lyrichu/" target="_blank" class="search-title"><i class="icon-quo-left icon"></i>博客园</a>
853            </li>
854          
855        </ul>
856  		
857    	</section>
858    
859
860    
861    	<section class="tools-section tools-section-me" q-show="aboutme">
862  	  	
863  	  		<div class="aboutme-wrap" id="js-aboutme">&lt;h1&gt;做一个热爱生活的程序员!&lt;/h1&gt;</div>
864  	  	
865    	</section>
866    
867  </div>
868  
869</div>
870    <!-- Root element of PhotoSwipe. Must have class pswp. -->
871<div class="pswp" tabindex="-1" role="dialog" aria-hidden="true">
872
873    <!-- Background of PhotoSwipe. 
874         It's a separate element as animating opacity is faster than rgba(). -->
875    <div class="pswp__bg"></div>
876
877    <!-- Slides wrapper with overflow:hidden. -->
878    <div class="pswp__scroll-wrap">
879
880        <!-- Container that holds slides. 
881            PhotoSwipe keeps only 3 of them in the DOM to save memory.
882            Don't modify these 3 pswp__item elements, data is added later on. -->
883        <div class="pswp__container">
884            <div class="pswp__item"></div>
885            <div class="pswp__item"></div>
886            <div class="pswp__item"></div>
887        </div>
888
889        <!-- Default (PhotoSwipeUI_Default) interface on top of sliding area. Can be changed. -->
890        <div class="pswp__ui pswp__ui--hidden">
891
892            <div class="pswp__top-bar">
893
894                <!--  Controls are self-explanatory. Order can be changed. -->
895
896                <div class="pswp__counter"></div>
897
898                <button class="pswp__button pswp__button--close" title="Close (Esc)"></button>
899
900                <button class="pswp__button pswp__button--share" style="display:none" title="Share"></button>
901
902                <button class="pswp__button pswp__button--fs" title="Toggle fullscreen"></button>
903
904                <button class="pswp__button pswp__button--zoom" title="Zoom in/out"></button>
905
906                <!-- Preloader demo http://codepen.io/dimsemenov/pen/yyBWoR -->
907                <!-- element will get class pswp__preloader--active when preloader is running -->
908                <div class="pswp__preloader">
909                    <div class="pswp__preloader__icn">
910                      <div class="pswp__preloader__cut">
911                        <div class="pswp__preloader__donut"></div>
912                      </div>
913                    </div>
914                </div>
915            </div>
916
917            <div class="pswp__share-modal pswp__share-modal--hidden pswp__single-tap">
918                <div class="pswp__share-tooltip"></div> 
919            </div>
920
921            <button class="pswp__button pswp__button--arrow--left" title="Previous (arrow left)">
922            </button>
923
924            <button class="pswp__button pswp__button--arrow--right" title="Next (arrow right)">
925            </button>
926
927            <div class="pswp__caption">
928                <div class="pswp__caption__center"></div>
929            </div>
930
931        </div>
932
933    </div>
934
935</div>
936  </div>
937</body>
938</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.