Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
切换分支/标签
查找文件 普通视图历史永久链接Permalink
new_layer.html 54.8 KB
Newer Older
T
Deploy to GitHub Pages: 98d9b34b2fee632d41115333583e1e9cbe43c738  
Travis CI 已提交
1 2
Y
Update docs  
Yu Yang 已提交
3 4 5 6 7 8 9 10 
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN"
  "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">


<html xmlns="http://www.w3.org/1999/xhtml">
  <head>
    <meta http-equiv="Content-Type" content="text/html; charset=utf-8" />
T
Deploy to GitHub Pages: 703cce35a871c78a5e3f1c33abf7fe788e2bbda5  
Travis CI 已提交
11 
    <title>Writing New Layers &#8212; PaddlePaddle  documentation</title>
Y
Update docs  
Yu Yang 已提交
12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 
    
    <link rel="stylesheet" href="../../_static/classic.css" type="text/css" />
    <link rel="stylesheet" href="../../_static/pygments.css" type="text/css" />
    
    <script type="text/javascript">
      var DOCUMENTATION_OPTIONS = {
        URL_ROOT:    '../../',
        VERSION:     '',
        COLLAPSE_INDEX: false,
        FILE_SUFFIX: '.html',
        HAS_SOURCE:  true
      };
    </script>
    <script type="text/javascript" src="../../_static/jquery.js"></script>
    <script type="text/javascript" src="../../_static/underscore.js"></script>
    <script type="text/javascript" src="../../_static/doctools.js"></script>
    <script type="text/javascript" src="https://cdn.mathjax.org/mathjax/latest/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script>
    <link rel="top" title="PaddlePaddle  documentation" href="../../index.html" />
Y
Upload docs  
Yu Yang 已提交
30 31 32 
    <link rel="up" title="Writing New Layers" href="index.html" />
    <link rel="next" title="Source Code Documents" href="../../source/index.html" />
    <link rel="prev" title="Writing New Layers" href="index.html" />
T
Deploy to GitHub Pages: 98d9b34b2fee632d41115333583e1e9cbe43c738  
Travis CI 已提交
33 34 35 36 37 38 39 40 41 42 
<script>
var _hmt = _hmt || [];
(function() {
  var hm = document.createElement("script");
  hm.src = "//hm.baidu.com/hm.js?b9a314ab40d04d805655aab1deee08ba";
  var s = document.getElementsByTagName("script")[0]; 
  s.parentNode.insertBefore(hm, s);
})();
</script>
Y
Update docs  
Yu Yang 已提交
43 44 45 46 47 48 49 50 51 52 53 54 
  </head>
  <body role="document">
    <div class="related" role="navigation" aria-label="related navigation">
      <h3>Navigation</h3>
      <ul>
        <li class="right" style="margin-right: 10px">
          <a href="../../genindex.html" title="General Index"
             accesskey="I">index</a></li>
        <li class="right" >
          <a href="../../py-modindex.html" title="Python Module Index"
             >modules</a> |</li>
        <li class="right" >
Y
Upload docs  
Yu Yang 已提交
55 
          <a href="../../source/index.html" title="Source Code Documents"
Y
Update docs  
Yu Yang 已提交
56 57 
             accesskey="N">next</a> |</li>
        <li class="right" >
Y
Upload docs  
Yu Yang 已提交
58 
          <a href="index.html" title="Writing New Layers"
Y
Update docs  
Yu Yang 已提交
59 
             accesskey="P">previous</a> |</li>
T
Deploy to GitHub Pages: 703cce35a871c78a5e3f1c33abf7fe788e2bbda5  
Travis CI 已提交
60 61 
        <li class="nav-item nav-item-0"><a href="../../index.html">PaddlePaddle  documentation</a> &#187;</li>
          <li class="nav-item nav-item-1"><a href="index.html" accesskey="U">Writing New Layers</a> &#187;</li>
Y
Update docs  
Yu Yang 已提交
62 63 64 65 66 67 68 69 70 
      </ul>
    </div>  

    <div class="document">
      <div class="documentwrapper">
        <div class="bodywrapper">
          <div class="body" role="main">
            
  <div class="section" id="writing-new-layers">
Y
Upload docs  
Yu Yang 已提交
71 
<h1>Writing New Layers<a class="headerlink" href="#writing-new-layers" title="Permalink to this headline"></a></h1>
Y
Update docs  
Yu Yang 已提交
72 73 74 75 
<p>This tutorial will guide you to write customized layers in PaddlePaddle. We will utilize fully connected layer as an example to guide you through the following steps for writing a new layer.</p>
<ul class="simple">
<li>Derive equations for the forward and backward part of the layer.</li>
<li>Implement C++ class for the layer.</li>
Y
Upload docs  
Yu Yang 已提交
76 77 
<li>Write gradient check unit test to make sure the gradients are correctly computed.</li>
<li>Implement Python wrapper for the layer.</li>
Y
Update docs  
Yu Yang 已提交
78 79 80 
</ul>
</div>
<div class="section" id="derive-equations">
Y
Upload docs  
Yu Yang 已提交
81 
<h1>Derive Equations<a class="headerlink" href="#derive-equations" title="Permalink to this headline"></a></h1>
Y
Update docs  
Yu Yang 已提交
82 
<p>First we need to derive equations of the <em>forward</em> and <em>backward</em> part of the layer. The forward part computes the output given an input. The backward part computes the gradients of the input and the parameters given the the gradients of the output.</p>
Y
Upload docs  
Yu Yang 已提交
83 84 
<p>The illustration of a fully connected layer is shown in the following figure. In a fully connected layer, all output nodes are connected to all the input nodes.</p>
<a class="reference internal image-reference" href="../../_images/FullyConnected.jpg"><img alt="../../_images/FullyConnected.jpg" class="align-center" src="../../_images/FullyConnected.jpg" style="width: 391.2px; height: 175.2px;" /></a>
Y
Update docs  
Yu Yang 已提交
85 
<p>The <em>forward part</em> of a layer transforms an input into the corresponding output.
Y
Upload docs  
Yu Yang 已提交
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 
Fully connected layer takes a dense input vector with dimension <span class="math">\(D_i\)</span>. It uses a transformation matrix <span class="math">\(W\)</span> with size <span class="math">\(D_i \times D_o\)</span> to project <span class="math">\(x\)</span> into a <span class="math">\(D_o\)</span> dimensional vector, and add a bias vector <span class="math">\(b\)</span> with dimension <span class="math">\(D_o\)</span> to the vector.</p>
<div class="math">
\[y = f(W^T x + b)\]</div>
<p>where <span class="math">\(f(.)\)</span> is an nonlinear <em>activation</em> function, such as sigmoid, tanh, and Relu.</p>
<p>The transformation matrix <span class="math">\(W\)</span> and bias vector <span class="math">\(b\)</span> are the <em>parameters</em> of the layer. The <em>parameters</em> of a layer are learned during training in the <em>backward pass</em>. The backward pass computes the gradients of the output function with respect to all parameters and inputs. The optimizer can use chain rule to compute the gradients of the loss function with respect to each parameter.</p>
<p>Suppose our loss function is <span class="math">\(c(y)\)</span>, then</p>
<div class="math">
\[\frac{\partial c(y)}{\partial x} = \frac{\partial c(y)}{\partial y} \frac{\partial y}{\partial x}\]</div>
<p>Suppose <span class="math">\(z = f(W^T x + b)\)</span>, then</p>
<div class="math">
\[\frac{\partial y}{\partial z} = \frac{\partial f(z)}{\partial z}\]</div>
<p>This derivative can be automatically computed by our base layer class.</p>
<p>Then, for fully connected layer, we need to compute:</p>
<div class="math">
\[\frac{\partial z}{\partial x} = W, \frac{\partial z_j}{\partial W_{ij}} = x_i, \frac{\partial z}{\partial b} = \mathbf 1\]</div>
<p>where <span class="math">\(\mathbf 1\)</span> is an all one vector, <span class="math">\(W_{ij}\)</span> is the number at the i-th row and j-th column of the matrix <span class="math">\(W\)</span>, <span class="math">\(z_j\)</span> is the j-th component of the vector <span class="math">\(z\)</span>, and <span class="math">\(x_i\)</span> is the i-th component of the vector <span class="math">\(x\)</span>.</p>
<p>Finally we can use chain rule to calculate <span class="math">\(\frac{\partial z}{\partial x}\)</span>, and <span class="math">\(\frac{\partial z}{\partial W}\)</span>. The details of the computation will be given in the next section.</p>
Y
Update docs  
Yu Yang 已提交
103 104 
</div>
<div class="section" id="implement-c-class">
Y
Upload docs  
Yu Yang 已提交
105 106 107 
<h1>Implement C++ Class<a class="headerlink" href="#implement-c-class" title="Permalink to this headline"></a></h1>
<p>The C++ class of the layer implements the initialization, forward, and backward part of the layer. The fully connected layer is at <code class="code docutils literal"><span class="pre">paddle/gserver/layers/FullyConnectedLayer.h</span></code> and <code class="code docutils literal"><span class="pre">paddle/gserver/layers/FullyConnectedLayer.cpp</span></code>. We list simplified version of the code below.</p>
<p>It needs to derive the base class <code class="code docutils literal"><span class="pre">paddle::BaseLayer</span></code>, and it needs to override the following functions:</p>
Y
Update docs  
Yu Yang 已提交
108 109 
<ul class="simple">
<li>constructor and destructor.</li>
Y
Upload docs  
Yu Yang 已提交
110 111 112 113 
<li><code class="code docutils literal"><span class="pre">init</span></code> function. It is used to initialize the parameters and settings.</li>
<li><code class="code docutils literal"><span class="pre">forward</span></code>. It implements the forward part of the layer.</li>
<li><code class="code docutils literal"><span class="pre">backward</span></code>. It implements the backward part of the layer.</li>
<li><code class="code docutils literal"><span class="pre">prefetch</span></code>. It is utilized to determine the rows corresponding parameter matrix to prefetch from parameter server. You do not need to override this function if your layer does not need remote sparse update. (most layers do not need to support remote sparse update)</li>
Y
Update docs  
Yu Yang 已提交
114 115 
</ul>
<p>The header file is listed below:</p>
Y
Upload docs  
Yu Yang 已提交
116 
<div class="highlight-c++"><div class="highlight"><pre><span></span><span class="k">namespace</span> <span class="n">paddle</span> <span class="p">{</span>
Y
Update docs  
Yu Yang 已提交
117 118 119 120 121 122 123 124 
<span class="cm">/**</span>
<span class="cm"> * A layer has full connections to all neurons in the previous layer.</span>
<span class="cm"> * It computes an inner product with a set of learned weights, and</span>
<span class="cm"> * (optionally) adds biases.</span>
<span class="cm"> *</span>
<span class="cm"> * The config file api is fc_layer.</span>
<span class="cm"> */</span>
Y
Upload docs  
Yu Yang 已提交
125 126 
<span class="k">class</span> <span class="nc">FullyConnectedLayer</span> <span class="o">:</span> <span class="k">public</span> <span class="n">Layer</span> <span class="p">{</span>
<span class="k">protected</span><span class="o">:</span>
Y
Update docs  
Yu Yang 已提交
127 128 129 
  <span class="n">WeightList</span> <span class="n">weights_</span><span class="p">;</span>
  <span class="n">std</span><span class="o">::</span><span class="n">unique_ptr</span><span class="o">&lt;</span><span class="n">Weight</span><span class="o">&gt;</span> <span class="n">biases_</span><span class="p">;</span>
Y
Upload docs  
Yu Yang 已提交
130 131 
<span class="k">public</span><span class="o">:</span>
  <span class="k">explicit</span> <span class="n">FullyConnectedLayer</span><span class="p">(</span><span class="k">const</span> <span class="n">LayerConfig</span><span class="o">&amp;</span> <span class="n">config</span><span class="p">)</span>
Y
Update docs  
Yu Yang 已提交
132 133 134 135 136 137 138 139 140 
      <span class="o">:</span> <span class="n">Layer</span><span class="p">(</span><span class="n">config</span><span class="p">)</span> <span class="p">{}</span>
  <span class="o">~</span><span class="n">FullyConnectedLayer</span><span class="p">()</span> <span class="p">{}</span>

  <span class="kt">bool</span> <span class="n">init</span><span class="p">(</span><span class="k">const</span> <span class="n">LayerMap</span><span class="o">&amp;</span> <span class="n">layerMap</span><span class="p">,</span> <span class="k">const</span> <span class="n">ParameterMap</span><span class="o">&amp;</span> <span class="n">parameterMap</span><span class="p">);</span>

  <span class="n">Weight</span><span class="o">&amp;</span> <span class="n">getWeight</span><span class="p">(</span><span class="kt">int</span> <span class="n">idx</span><span class="p">)</span> <span class="p">{</span> <span class="k">return</span> <span class="o">*</span><span class="n">weights_</span><span class="p">[</span><span class="n">idx</span><span class="p">];</span> <span class="p">}</span>

  <span class="kt">void</span> <span class="n">prefetch</span><span class="p">();</span>
  <span class="kt">void</span> <span class="nf">forward</span><span class="p">(</span><span class="n">PassType</span> <span class="n">passType</span><span class="p">);</span>
Y
Upload docs  
Yu Yang 已提交
141 
  <span class="kt">void</span> <span class="nf">backward</span><span class="p">(</span><span class="k">const</span> <span class="n">UpdateCallback</span><span class="o">&amp;</span> <span class="n">callback</span> <span class="o">=</span> <span class="k">nullptr</span><span class="p">);</span>
Y
Update docs  
Yu Yang 已提交
142 143 144 145 
<span class="p">};</span>
<span class="p">}</span>  <span class="c1">// namespace paddle</span>
</pre></div>
</div>
Y
Upload docs  
Yu Yang 已提交
146 
<p>It defines the parameters as class variables. We use <code class="code docutils literal"><span class="pre">Weight</span></code> class as abstraction of parameters. It supports multi-thread update. The details of this class will be described in details in the implementations.</p>
Y
Update docs  
Yu Yang 已提交
147 
<ul class="simple">
Y
Upload docs  
Yu Yang 已提交
148 149 
<li><code class="code docutils literal"><span class="pre">weights_</span></code> is a list of weights for the transformation matrices. The current implementation can have more than one inputs. Thus, it has a list of weights. One weight corresponds to an input.</li>
<li><code class="code docutils literal"><span class="pre">biases_</span></code> is a weight for the bias vector.</li>
Y
Update docs  
Yu Yang 已提交
150 
</ul>
Y
Upload docs  
Yu Yang 已提交
151 152 
<p>The fully connected layer does not have layer configuration hyper-parameters. If there are some layer hyper-parameters, a common practice is to store it in <code class="code docutils literal"><span class="pre">LayerConfig&amp;</span> <span class="pre">config</span></code>, and put it into a class variable in the constructor.</p>
<p>The following code snippet implements the <code class="code docutils literal"><span class="pre">init</span></code> function.</p>
Y
Update docs  
Yu Yang 已提交
153 
<ul class="simple">
Y
Upload docs  
Yu Yang 已提交
154 155 
<li>First, every <code class="code docutils literal"><span class="pre">init</span></code> function must call the <code class="code docutils literal"><span class="pre">init</span></code> function of the base class <code class="code docutils literal"><span class="pre">Layer::init(layerMap,</span> <span class="pre">parameterMap);</span></code>. This statement will initialize the required variables and connections for each layer.</li>
<li>The it initializes all the weights matrices <span class="math">\(W\)</span>. The current implementation can have more than one inputs. Thus, it has a list of weights.</li>
Y
Update docs  
Yu Yang 已提交
156 157 
<li>Finally, it initializes the bias.</li>
</ul>
Y
Upload docs  
Yu Yang 已提交
158 
<div class="highlight-c++"><div class="highlight"><pre><span></span><span class="kt">bool</span> <span class="n">FullyConnectedLayer</span><span class="o">::</span><span class="n">init</span><span class="p">(</span><span class="k">const</span> <span class="n">LayerMap</span><span class="o">&amp;</span> <span class="n">layerMap</span><span class="p">,</span>
Y
Update docs  
Yu Yang 已提交
159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 
                               <span class="k">const</span> <span class="n">ParameterMap</span><span class="o">&amp;</span> <span class="n">parameterMap</span><span class="p">)</span> <span class="p">{</span>
  <span class="cm">/* Initialize the basic parent class */</span>
  <span class="n">Layer</span><span class="o">::</span><span class="n">init</span><span class="p">(</span><span class="n">layerMap</span><span class="p">,</span> <span class="n">parameterMap</span><span class="p">);</span>

  <span class="cm">/* initialize the weightList */</span>
  <span class="n">CHECK</span><span class="p">(</span><span class="n">inputLayers_</span><span class="p">.</span><span class="n">size</span><span class="p">()</span> <span class="o">==</span> <span class="n">parameters_</span><span class="p">.</span><span class="n">size</span><span class="p">());</span>
  <span class="k">for</span> <span class="p">(</span><span class="kt">size_t</span> <span class="n">i</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">i</span> <span class="o">&lt;</span> <span class="n">inputLayers_</span><span class="p">.</span><span class="n">size</span><span class="p">();</span> <span class="n">i</span><span class="o">++</span><span class="p">)</span> <span class="p">{</span>
    <span class="c1">// Option the parameters</span>
    <span class="kt">size_t</span> <span class="n">height</span> <span class="o">=</span> <span class="n">inputLayers_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">getSize</span><span class="p">();</span>
    <span class="kt">size_t</span> <span class="n">width</span> <span class="o">=</span> <span class="n">getSize</span><span class="p">();</span>

    <span class="c1">// create a new weight</span>
    <span class="k">if</span> <span class="p">(</span><span class="n">parameters_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">isSparse</span><span class="p">())</span> <span class="p">{</span>
      <span class="n">CHECK_LE</span><span class="p">(</span><span class="n">parameters_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">getSize</span><span class="p">(),</span> <span class="n">width</span> <span class="o">*</span> <span class="n">height</span><span class="p">);</span>
    <span class="p">}</span> <span class="k">else</span> <span class="p">{</span>
      <span class="n">CHECK_EQ</span><span class="p">(</span><span class="n">parameters_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">getSize</span><span class="p">(),</span> <span class="n">width</span> <span class="o">*</span> <span class="n">height</span><span class="p">);</span>
    <span class="p">}</span>
Y
Upload docs  
Yu Yang 已提交
176 
    <span class="n">Weight</span><span class="o">*</span> <span class="n">w</span> <span class="o">=</span> <span class="k">new</span> <span class="n">Weight</span><span class="p">(</span><span class="n">height</span><span class="p">,</span> <span class="n">width</span><span class="p">,</span> <span class="n">parameters_</span><span class="p">[</span><span class="n">i</span><span class="p">]);</span>
Y
Update docs  
Yu Yang 已提交
177 178 179 180 181 182 183 

    <span class="c1">// append the new weight to the list</span>
    <span class="n">weights_</span><span class="p">.</span><span class="n">emplace_back</span><span class="p">(</span><span class="n">w</span><span class="p">);</span>
  <span class="p">}</span>

  <span class="cm">/* initialize biases_ */</span>
  <span class="k">if</span> <span class="p">(</span><span class="n">biasParameter_</span><span class="p">.</span><span class="n">get</span><span class="p">()</span> <span class="o">!=</span> <span class="nb">NULL</span><span class="p">)</span> <span class="p">{</span>
Y
Upload docs  
Yu Yang 已提交
184 
    <span class="n">biases_</span> <span class="o">=</span> <span class="n">std</span><span class="o">::</span><span class="n">unique_ptr</span><span class="o">&lt;</span><span class="n">Weight</span><span class="o">&gt;</span><span class="p">(</span><span class="k">new</span> <span class="n">Weight</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">getSize</span><span class="p">(),</span> <span class="n">biasParameter_</span><span class="p">));</span>
Y
Update docs  
Yu Yang 已提交
185 186 187 188 189 190 191 192 
  <span class="p">}</span>

  <span class="k">return</span> <span class="nb">true</span><span class="p">;</span>
<span class="p">}</span>
</pre></div>
</div>
<p>The implementation of the forward part has the following steps.</p>
<ul class="simple">
Y
Upload docs  
Yu Yang 已提交
193 194 195 196 
<li>Every layer must call <code class="code docutils literal"><span class="pre">Layer::forward(passType);</span></code> at the beginning of its <code class="code docutils literal"><span class="pre">forward</span></code> function.</li>
<li>Then it allocates memory for the output using <code class="code docutils literal"><span class="pre">reserveOutput(batchSize,</span> <span class="pre">size);</span></code>. This step is necessary because we support the batches to have different batch sizes. <code class="code docutils literal"><span class="pre">reserveOutput</span></code> will change the size of the output accordingly. For the sake of efficiency, we will allocate new memory if we want to expand the matrix, but we will reuse the existing memory block if we want to shrink the matrix.</li>
<li>Then it computes <span class="math">\(\sum_i W_i x + b\)</span> using Matrix operations. <code class="code docutils literal"><span class="pre">getInput(i).value</span></code> retrieve the matrix of the i-th input. Each input is a <span class="math">\(batchSize \times dim\)</span> matrix, where each row represents an single input in a batch. For a complete lists of supported matrix operations, please refer to <code class="code docutils literal"><span class="pre">paddle/math/Matrix.h</span></code> and <code class="code docutils literal"><span class="pre">paddle/math/BaseMatrix.h</span></code>.</li>
<li>Finally it applies the activation function using <code class="code docutils literal"><span class="pre">forwardActivation();</span></code>. It will automatically applies the corresponding activation function specifies in the network configuration.</li>
Y
Update docs  
Yu Yang 已提交
197 
</ul>
Y
Upload docs  
Yu Yang 已提交
198 
<div class="highlight-c++"><div class="highlight"><pre><span></span><span class="kt">void</span> <span class="n">FullyConnectedLayer</span><span class="o">::</span><span class="n">forward</span><span class="p">(</span><span class="n">PassType</span> <span class="n">passType</span><span class="p">)</span> <span class="p">{</span>
Y
Update docs  
Yu Yang 已提交
199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 
  <span class="n">Layer</span><span class="o">::</span><span class="n">forward</span><span class="p">(</span><span class="n">passType</span><span class="p">);</span>

  <span class="cm">/* malloc memory for the output_ if necessary */</span>
  <span class="kt">int</span> <span class="n">batchSize</span> <span class="o">=</span> <span class="n">getInput</span><span class="p">(</span><span class="mi">0</span><span class="p">).</span><span class="n">getBatchSize</span><span class="p">();</span>
  <span class="kt">int</span> <span class="n">size</span> <span class="o">=</span> <span class="n">getSize</span><span class="p">();</span>

  <span class="p">{</span>
    <span class="c1">// Settup the size of the output.</span>
    <span class="n">reserveOutput</span><span class="p">(</span><span class="n">batchSize</span><span class="p">,</span> <span class="n">size</span><span class="p">);</span>
  <span class="p">}</span>

  <span class="n">MatrixPtr</span> <span class="n">outV</span> <span class="o">=</span> <span class="n">getOutputValue</span><span class="p">();</span>

  <span class="c1">// Apply the the transformation matrix to each input.</span>
  <span class="k">for</span> <span class="p">(</span><span class="kt">size_t</span> <span class="n">i</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">i</span> <span class="o">!=</span> <span class="n">inputLayers_</span><span class="p">.</span><span class="n">size</span><span class="p">();</span> <span class="o">++</span><span class="n">i</span><span class="p">)</span> <span class="p">{</span>
    <span class="k">auto</span> <span class="n">input</span> <span class="o">=</span> <span class="n">getInput</span><span class="p">(</span><span class="n">i</span><span class="p">);</span>
    <span class="n">CHECK</span><span class="p">(</span><span class="n">input</span><span class="p">.</span><span class="n">value</span><span class="p">)</span> <span class="o">&lt;&lt;</span> <span class="s">&quot;The input of &#39;fc&#39; layer must be matrix&quot;</span><span class="p">;</span>
    <span class="n">i</span> <span class="o">==</span> <span class="mi">0</span> <span class="o">?</span> <span class="n">outV</span><span class="o">-&gt;</span><span class="n">mul</span><span class="p">(</span><span class="n">input</span><span class="p">.</span><span class="n">value</span><span class="p">,</span> <span class="n">weights_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">getW</span><span class="p">(),</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
           <span class="o">:</span> <span class="n">outV</span><span class="o">-&gt;</span><span class="n">mul</span><span class="p">(</span><span class="n">input</span><span class="p">.</span><span class="n">value</span><span class="p">,</span> <span class="n">weights_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">getW</span><span class="p">(),</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">);</span>
  <span class="p">}</span>

  <span class="cm">/* add the bias-vector */</span>
  <span class="k">if</span> <span class="p">(</span><span class="n">biases_</span><span class="p">.</span><span class="n">get</span><span class="p">()</span> <span class="o">!=</span> <span class="nb">NULL</span><span class="p">)</span> <span class="p">{</span>
    <span class="n">outV</span><span class="o">-&gt;</span><span class="n">addBias</span><span class="p">(</span><span class="o">*</span><span class="p">(</span><span class="n">biases_</span><span class="o">-&gt;</span><span class="n">getW</span><span class="p">()),</span> <span class="mi">1</span><span class="p">);</span>
  <span class="p">}</span>

  <span class="cm">/* activation */</span> <span class="p">{</span>
    <span class="n">forwardActivation</span><span class="p">();</span>
  <span class="p">}</span>
<span class="p">}</span>
</pre></div>
</div>
<p>The implementation of the backward part has the following steps.</p>
<ul class="simple">
Y
Upload docs  
Yu Yang 已提交
233 234 235 
<li><code class="code docutils literal"><span class="pre">backwardActivation()</span></code> computes the gradients of the activation. The gradients will be multiplies in place to the gradients of the output, which can be retrieved using <code class="code docutils literal"><span class="pre">getOutputGrad()</span></code>.</li>
<li>Compute the gradients of bias. Notice that we an use <code class="code docutils literal"><span class="pre">biases_-&gt;getWGrad()</span></code> to get the gradient matrix of the corresponding parameter. After the gradient of one parameter is updated, it <strong>MUST</strong> call <code class="code docutils literal"><span class="pre">getParameterPtr()-&gt;incUpdate(callback);</span></code>. This is utilize for parameter update over multiple threads or multiple machines.</li>
<li>Then it computes the gradients of the transformation matrices and inputs, and it calls <code class="code docutils literal"><span class="pre">incUpdate</span></code> for the corresponding parameter. This gives the framework the chance to know whether it has gathered all the gradient to one parameter so that it can do some overlapping work (e.g., network communication)</li>
Y
Update docs  
Yu Yang 已提交
236 
</ul>
Y
Upload docs  
Yu Yang 已提交
237 
<div class="highlight-c++"><div class="highlight"><pre><span></span><span class="kt">void</span> <span class="n">FullyConnectedLayer</span><span class="o">::</span><span class="n">backward</span><span class="p">(</span><span class="k">const</span> <span class="n">UpdateCallback</span><span class="o">&amp;</span> <span class="n">callback</span><span class="p">)</span> <span class="p">{</span>
Y
Update docs  
Yu Yang 已提交
238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 
  <span class="cm">/* Do derivation for activations.*/</span> <span class="p">{</span>
    <span class="n">backwardActivation</span><span class="p">();</span>
  <span class="p">}</span>

  <span class="k">if</span> <span class="p">(</span><span class="n">biases_</span> <span class="o">&amp;&amp;</span> <span class="n">biases_</span><span class="o">-&gt;</span><span class="n">getWGrad</span><span class="p">())</span> <span class="p">{</span>
    <span class="n">biases_</span><span class="o">-&gt;</span><span class="n">getWGrad</span><span class="p">()</span><span class="o">-&gt;</span><span class="n">collectBias</span><span class="p">(</span><span class="o">*</span><span class="n">getOutputGrad</span><span class="p">(),</span> <span class="mi">1</span><span class="p">);</span>

    <span class="cm">/* Increasing the number of gradient */</span>
    <span class="n">biases_</span><span class="o">-&gt;</span><span class="n">getParameterPtr</span><span class="p">()</span><span class="o">-&gt;</span><span class="n">incUpdate</span><span class="p">(</span><span class="n">callback</span><span class="p">);</span>
  <span class="p">}</span>

  <span class="kt">bool</span> <span class="n">syncFlag</span> <span class="o">=</span> <span class="n">hl_get_sync_flag</span><span class="p">();</span>

  <span class="k">for</span> <span class="p">(</span><span class="kt">size_t</span> <span class="n">i</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">i</span> <span class="o">!=</span> <span class="n">inputLayers_</span><span class="p">.</span><span class="n">size</span><span class="p">();</span> <span class="o">++</span><span class="n">i</span><span class="p">)</span> <span class="p">{</span>
    <span class="cm">/* Calculate the W-gradient for the current layer */</span>
    <span class="k">if</span> <span class="p">(</span><span class="n">weights_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">getWGrad</span><span class="p">())</span> <span class="p">{</span>
      <span class="n">MatrixPtr</span> <span class="n">input_T</span> <span class="o">=</span> <span class="n">getInputValue</span><span class="p">(</span><span class="n">i</span><span class="p">)</span><span class="o">-&gt;</span><span class="n">getTranspose</span><span class="p">();</span>
      <span class="n">MatrixPtr</span> <span class="n">oGrad</span> <span class="o">=</span> <span class="n">getOutputGrad</span><span class="p">();</span>
      <span class="p">{</span>
        <span class="n">weights_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">getWGrad</span><span class="p">()</span><span class="o">-&gt;</span><span class="n">mul</span><span class="p">(</span><span class="n">input_T</span><span class="p">,</span> <span class="n">oGrad</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">);</span>
      <span class="p">}</span>
    <span class="p">}</span>


    <span class="cm">/* Calculate the input layers error */</span>
    <span class="n">MatrixPtr</span> <span class="n">preGrad</span> <span class="o">=</span> <span class="n">getInputGrad</span><span class="p">(</span><span class="n">i</span><span class="p">);</span>
    <span class="k">if</span> <span class="p">(</span><span class="nb">NULL</span> <span class="o">!=</span> <span class="n">preGrad</span><span class="p">)</span> <span class="p">{</span>
      <span class="n">MatrixPtr</span> <span class="n">weights_T</span> <span class="o">=</span> <span class="n">weights_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">getW</span><span class="p">()</span><span class="o">-&gt;</span><span class="n">getTranspose</span><span class="p">();</span>
      <span class="n">preGrad</span><span class="o">-&gt;</span><span class="n">mul</span><span class="p">(</span><span class="n">getOutputGrad</span><span class="p">(),</span> <span class="n">weights_T</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">);</span>
    <span class="p">}</span>

    <span class="p">{</span>
      <span class="n">weights_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">getParameterPtr</span><span class="p">()</span><span class="o">-&gt;</span><span class="n">incUpdate</span><span class="p">(</span><span class="n">callback</span><span class="p">);</span>
    <span class="p">}</span>
  <span class="p">}</span>
<span class="p">}</span>
</pre></div>
</div>
Y
Upload docs  
Yu Yang 已提交
276 
<p>The <code class="code docutils literal"><span class="pre">prefetch</span></code> function specifies the rows that need to be fetched from parameter server during training. It is only useful for remote sparse training. In remote sparse training, the full parameter matrix is stored distributedly at the parameter server. When the layer uses a batch for training, only a subset of locations of the input is non-zero in this batch. Thus, this layer only needs the rows of the transformation matrix corresponding to the locations of these non-zero entries. The <code class="code docutils literal"><span class="pre">prefetch</span></code> function specifies the ids of these rows.</p>
Y
Update docs  
Yu Yang 已提交
277 
<p>Most of the layers do not need remote sparse training function. You do not need to override this function in this case.</p>
Y
Upload docs  
Yu Yang 已提交
278 
<div class="highlight-c++"><div class="highlight"><pre><span></span><span class="kt">void</span> <span class="n">FullyConnectedLayer</span><span class="o">::</span><span class="n">prefetch</span><span class="p">()</span> <span class="p">{</span>
Y
Update docs  
Yu Yang 已提交
279 280 
  <span class="k">for</span> <span class="p">(</span><span class="kt">size_t</span> <span class="n">i</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">i</span> <span class="o">!=</span> <span class="n">inputLayers_</span><span class="p">.</span><span class="n">size</span><span class="p">();</span> <span class="o">++</span><span class="n">i</span><span class="p">)</span> <span class="p">{</span>
    <span class="k">auto</span><span class="o">*</span> <span class="n">sparseParam</span> <span class="o">=</span>
Y
Upload docs  
Yu Yang 已提交
281 
        <span class="k">dynamic_cast</span><span class="o">&lt;</span><span class="n">SparsePrefetchRowCpuMatrix</span><span class="o">*&gt;</span><span class="p">(</span><span class="n">weights_</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">-&gt;</span><span class="n">getW</span><span class="p">().</span><span class="n">get</span><span class="p">());</span>
Y
Update docs  
Yu Yang 已提交
282 283 284 285 286 287 288 289 
    <span class="k">if</span> <span class="p">(</span><span class="n">sparseParam</span><span class="p">)</span> <span class="p">{</span>
      <span class="n">MatrixPtr</span> <span class="n">input</span> <span class="o">=</span> <span class="n">getInputValue</span><span class="p">(</span><span class="n">i</span><span class="p">);</span>
      <span class="n">sparseParam</span><span class="o">-&gt;</span><span class="n">addRows</span><span class="p">(</span><span class="n">input</span><span class="p">);</span>
    <span class="p">}</span>
  <span class="p">}</span>
<span class="p">}</span>
</pre></div>
</div>
Y
Upload docs  
Yu Yang 已提交
290 291 
<p>Finally, you can use <code class="code docutils literal"><span class="pre">REGISTER_LAYER(fc,</span> <span class="pre">FullyConnectedLayer);</span></code> to register the layer. <code class="code docutils literal"><span class="pre">fc</span></code> is the identifier of the layer, and <code class="code docutils literal"><span class="pre">FullyConnectedLayer</span></code> is the class name of the layer.</p>
<div class="highlight-c++"><div class="highlight"><pre><span></span><span class="k">namespace</span> <span class="n">paddle</span> <span class="p">{</span>
Y
Update docs  
Yu Yang 已提交
292 293 294 295 
<span class="n">REGISTER_LAYER</span><span class="p">(</span><span class="n">fc</span><span class="p">,</span> <span class="n">FullyConnectedLayer</span><span class="p">);</span>
<span class="p">}</span>
</pre></div>
</div>
Y
Upload docs  
Yu Yang 已提交
296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 
<p>If the <code class="code docutils literal"><span class="pre">cpp</span></code> file is put into <code class="code docutils literal"><span class="pre">paddle/gserver/layers</span></code>, it will be automatically added to the compilation list.</p>
</div>
<div class="section" id="write-gradient-check-unit-test">
<h1>Write Gradient Check Unit Test<a class="headerlink" href="#write-gradient-check-unit-test" title="Permalink to this headline"></a></h1>
<p>An easy way to verify the correctness of new layer&#8217;s implementation is to write a gradient check unit test. Gradient check unit test utilizes finite difference method to verify the gradient of a layer. It modifies the input with a small perturbation <span class="math">\(\Delta x\)</span> and observes the changes of output <span class="math">\(\Delta y\)</span>, the gradient can be computed as <span class="math">\(\frac{\Delta y}{\Delta x }\)</span>. This gradient can be compared with the gradient computed by the <code class="code docutils literal"><span class="pre">backward</span></code> function of the layer to ensure the correctness of the gradient computation. Notice that the gradient check only tests the correctness of the gradient computation, it does not necessarily guarantee the correctness of the implementation of the <code class="code docutils literal"><span class="pre">forward</span></code> and <code class="code docutils literal"><span class="pre">backward</span></code> function. You need to write more sophisticated unit tests to make sure your layer is implemented correctly.</p>
<p>All the gradient check unit tests are located in <code class="code docutils literal"><span class="pre">paddle/gserver/tests/test_LayerGrad.cpp</span></code>. You are recommended to put your test into a new test file if you are planning to write a new layer. The gradient test of the gradient check unit test of the fully connected layer is listed below. It has the following steps.</p>
<ul>
<li><dl class="first docutils">
<dt>Create layer configuration. A layer configuration can include the following attributes:</dt>
<dd><ul class="first last simple">
<li>size of the bias parameter. (4096 in our example)</li>
<li>type of the layer. (fc in our example)</li>
<li>size of the layer. (4096 in our example)</li>
<li>activation type. (softmax in our example)</li>
<li>dropout rate. (0.1 in our example)</li>
</ul>
</dd>
</dl>
</li>
<li><dl class="first docutils">
<dt>configure the input of the layer. In our example, we have only one input.</dt>
<dd><ul class="first last">
<li><dl class="first docutils">
<dt>type of the input (<code class="code docutils literal"><span class="pre">INPUT_DATA</span></code>) in our example. It can be one of the following types</dt>
<dd><ul class="first last simple">
<li><code class="code docutils literal"><span class="pre">INPUT_DATA</span></code>: dense vector.</li>
<li><code class="code docutils literal"><span class="pre">INPUT_LABEL</span></code>: integer.</li>
<li><code class="code docutils literal"><span class="pre">INPUT_DATA_TARGET</span></code>: dense vector, but it does not used to compute gradient.</li>
<li><code class="code docutils literal"><span class="pre">INPUT_SEQUENCE_DATA</span></code>: dense vector with sequence information.</li>
<li><code class="code docutils literal"><span class="pre">INPUT_HASSUB_SEQUENCE_DATA</span></code>: dense vector with both sequence and sub-sequence information.</li>
<li><code class="code docutils literal"><span class="pre">INPUT_SEQUENCE_LABEL</span></code>: integer with sequence information.</li>
<li><code class="code docutils literal"><span class="pre">INPUT_SPARSE_NON_VALUE_DATA</span></code>: 0-1 sparse data.</li>
<li><code class="code docutils literal"><span class="pre">INPUT_SPARSE_FLOAT_VALUE_DATA</span></code>: float sparse data.</li>
</ul>
</dd>
</dl>
</li>
<li><p class="first">name of the input. (<code class="code docutils literal"><span class="pre">layer_0</span></code> in our example)</p>
</li>
<li><p class="first">size of the input. (8192 in our example)</p>
</li>
<li><p class="first">number of non-zeros, only useful for sparse inputs.</p>
</li>
<li><p class="first">format of sparse data, only useful for sparse inputs.</p>
</li>
</ul>
</dd>
</dl>
</li>
<li><p class="first">each inputs needs to call <code class="code docutils literal"><span class="pre">config.layerConfig.add_inputs();</span></code> once.</p>
</li>
<li><dl class="first docutils">
<dt>call <code class="code docutils literal"><span class="pre">testLayerGrad</span></code> to perform gradient checks. It has the following arguments.</dt>
<dd><ul class="first last simple">
<li>layer and input configurations. (<code class="code docutils literal"><span class="pre">config</span></code> in our example)</li>
<li>type of the input. (<code class="code docutils literal"><span class="pre">fc</span></code> in our example)</li>
<li>batch size of the gradient check. (100 in our example)</li>
<li>whether the input is transpose. Most layers need to set it to <code class="code docutils literal"><span class="pre">false</span></code>. (<code class="code docutils literal"><span class="pre">false</span></code> in our example)</li>
<li>whether to use weights. Some layers or activations perform normalization so that the sum of their output is a constant. For example, the sum of output of a softmax activation is one. In this case, we cannot correctly compute the gradients using regular gradient check techniques. A weighted sum of the output, which is not a constant, is utilized to compute the gradients. (<code class="code docutils literal"><span class="pre">true</span></code> in our example, because the activation of a fully connected layer can be softmax)</li>
</ul>
</dd>
</dl>
</li>
</ul>
<div class="highlight-c++"><div class="highlight"><pre><span></span><span class="kt">void</span> <span class="nf">testFcLayer</span><span class="p">(</span><span class="n">string</span> <span class="n">format</span><span class="p">,</span> <span class="kt">size_t</span> <span class="n">nnz</span><span class="p">)</span> <span class="p">{</span>
  <span class="c1">// Create layer configuration.</span>
  <span class="n">TestConfig</span> <span class="n">config</span><span class="p">;</span>
  <span class="n">config</span><span class="p">.</span><span class="n">biasSize</span> <span class="o">=</span> <span class="mi">4096</span><span class="p">;</span>
  <span class="n">config</span><span class="p">.</span><span class="n">layerConfig</span><span class="p">.</span><span class="n">set_type</span><span class="p">(</span><span class="s">&quot;fc&quot;</span><span class="p">);</span>
  <span class="n">config</span><span class="p">.</span><span class="n">layerConfig</span><span class="p">.</span><span class="n">set_size</span><span class="p">(</span><span class="mi">4096</span><span class="p">);</span>
  <span class="n">config</span><span class="p">.</span><span class="n">layerConfig</span><span class="p">.</span><span class="n">set_active_type</span><span class="p">(</span><span class="s">&quot;sigmoid&quot;</span><span class="p">);</span>
  <span class="n">config</span><span class="p">.</span><span class="n">layerConfig</span><span class="p">.</span><span class="n">set_drop_rate</span><span class="p">(</span><span class="mf">0.1</span><span class="p">);</span>
  <span class="c1">// Setup inputs.</span>
  <span class="n">config</span><span class="p">.</span><span class="n">inputDefs</span><span class="p">.</span><span class="n">push_back</span><span class="p">(</span>
      <span class="p">{</span><span class="n">INPUT_DATA</span><span class="p">,</span> <span class="s">&quot;layer_0&quot;</span><span class="p">,</span> <span class="mi">8192</span><span class="p">,</span> <span class="n">nnz</span><span class="p">,</span> <span class="n">ParaSparse</span><span class="p">(</span><span class="n">format</span><span class="p">)});</span>
    <span class="n">config</span><span class="p">.</span><span class="n">layerConfig</span><span class="p">.</span><span class="n">add_inputs</span><span class="p">();</span>
  <span class="n">LOG</span><span class="p">(</span><span class="n">INFO</span><span class="p">)</span> <span class="o">&lt;&lt;</span> <span class="n">config</span><span class="p">.</span><span class="n">inputDefs</span><span class="p">[</span><span class="mi">0</span><span class="p">].</span><span class="n">sparse</span><span class="p">.</span><span class="n">sparse</span> <span class="o">&lt;&lt;</span> <span class="s">&quot; &quot;</span>
            <span class="o">&lt;&lt;</span> <span class="n">config</span><span class="p">.</span><span class="n">inputDefs</span><span class="p">[</span><span class="mi">0</span><span class="p">].</span><span class="n">sparse</span><span class="p">.</span><span class="n">format</span><span class="p">;</span>
  <span class="k">for</span> <span class="p">(</span><span class="k">auto</span> <span class="nl">useGpu</span> <span class="p">:</span> <span class="p">{</span><span class="nb">false</span><span class="p">,</span> <span class="nb">true</span><span class="p">})</span> <span class="p">{</span>
    <span class="n">testLayerGrad</span><span class="p">(</span><span class="n">config</span><span class="p">,</span> <span class="s">&quot;fc&quot;</span><span class="p">,</span> <span class="mi">100</span><span class="p">,</span> <span class="cm">/* trans */</span> <span class="nb">false</span><span class="p">,</span> <span class="n">useGpu</span><span class="p">,</span>
                  <span class="cm">/* weight */</span> <span class="nb">true</span><span class="p">);</span>
  <span class="p">}</span>
<span class="p">}</span>
</pre></div>
</div>
<p>If you are creating a new file for the test, such as <code class="code docutils literal"><span class="pre">paddle/gserver/tests/testFCGrad.cpp</span></code>, you need to add the file to <code class="code docutils literal"><span class="pre">paddle/gserver/tests/CMakeLists.txt</span></code>. An example is given below. All the unit tests will run when you execute the command <code class="code docutils literal"><span class="pre">make</span> <span class="pre">tests</span></code>. Notice that some layers might need high accuracy for the gradient check unit tests to work well. You need to configure <code class="code docutils literal"><span class="pre">WITH_DOUBLE</span></code> to <cite>ON</cite> when configuring cmake.</p>
<div class="highlight-bash"><div class="highlight"><pre><span></span>add_unittest_without_exec<span class="o">(</span>test_FCGrad
    test_FCGrad.cpp
    LayerGradUtil.cpp
    TestUtil.cpp<span class="o">)</span>

add_test<span class="o">(</span>NAME test_FCGrad
    COMMAND test_FCGrad<span class="o">)</span>
</pre></div>
</div>
Y
Update docs  
Yu Yang 已提交
391 392 
</div>
<div class="section" id="implement-python-wrapper">
Y
Upload docs  
Yu Yang 已提交
393 394 395 396 397 398 399 400 401 
<h1>Implement Python Wrapper<a class="headerlink" href="#implement-python-wrapper" title="Permalink to this headline"></a></h1>
<p>Implementing Python wrapper allows us to use the added layer in configuration files. All the Python wrappers are in file <code class="code docutils literal"><span class="pre">python/paddle/trainer/config_parser.py</span></code>. An example of the Python wrapper for fully connected layer is listed below. It has the following steps:</p>
<ul>
<li><p class="first">Use <code class="code docutils literal"><span class="pre">&#64;config_layer('fc')</span></code> at the decorator for all the Python wrapper class. <code class="code docutils literal"><span class="pre">fc</span></code> is the identifier of the layer.</p>
</li>
<li><dl class="first docutils">
<dt>Implements <code class="code docutils literal"><span class="pre">__init__</span></code> constructor function.</dt>
<dd><ul class="first last simple">
<li>It first call <code class="code docutils literal"><span class="pre">super(FCLayer,</span> <span class="pre">self).__init__(name,</span> <span class="pre">'fc',</span> <span class="pre">size,</span> <span class="pre">inputs=inputs,</span> <span class="pre">**xargs)</span></code> base constructor function. <code class="code docutils literal"><span class="pre">FCLayer</span></code> is the Python wrapper class name, and <code class="code docutils literal"><span class="pre">fc</span></code> is the layer identifier name. They must be correct in order for the wrapper to work.</li>
Y
Update docs  
Yu Yang 已提交
402 403 
<li>Then it computes the size and format (whether sparse) of each transformation matrix as well as the size.</li>
</ul>
Y
Upload docs  
Yu Yang 已提交
404 405 
</dd>
</dl>
Y
Update docs  
Yu Yang 已提交
406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 
</li>
</ul>
<div class="highlight-python"><div class="highlight"><pre><span></span><span class="nd">@config_layer</span><span class="p">(</span><span class="s1">&#39;fc&#39;</span><span class="p">)</span>
<span class="k">class</span> <span class="nc">FCLayer</span><span class="p">(</span><span class="n">LayerBase</span><span class="p">):</span>
    <span class="k">def</span> <span class="nf">__init__</span><span class="p">(</span>
            <span class="bp">self</span><span class="p">,</span>
            <span class="n">name</span><span class="p">,</span>
            <span class="n">size</span><span class="p">,</span>
            <span class="n">inputs</span><span class="p">,</span>
            <span class="n">bias</span><span class="o">=</span><span class="bp">True</span><span class="p">,</span>
            <span class="o">**</span><span class="n">xargs</span><span class="p">):</span>
        <span class="nb">super</span><span class="p">(</span><span class="n">FCLayer</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="n">__init__</span><span class="p">(</span><span class="n">name</span><span class="p">,</span> <span class="s1">&#39;fc&#39;</span><span class="p">,</span> <span class="n">size</span><span class="p">,</span> <span class="n">inputs</span><span class="o">=</span><span class="n">inputs</span><span class="p">,</span> <span class="o">**</span><span class="n">xargs</span><span class="p">)</span>
        <span class="k">for</span> <span class="n">input_index</span> <span class="ow">in</span> <span class="nb">xrange</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">inputs</span><span class="p">)):</span>
            <span class="n">input_layer</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">get_input_layer</span><span class="p">(</span><span class="n">input_index</span><span class="p">)</span>
            <span class="n">psize</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">size</span> <span class="o">*</span> <span class="n">input_layer</span><span class="o">.</span><span class="n">size</span>
            <span class="n">dims</span> <span class="o">=</span> <span class="p">[</span><span class="n">input_layer</span><span class="o">.</span><span class="n">size</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">size</span><span class="p">]</span>
            <span class="n">format</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">inputs</span><span class="p">[</span><span class="n">input_index</span><span class="p">]</span><span class="o">.</span><span class="n">format</span>
            <span class="n">sparse</span> <span class="o">=</span> <span class="n">format</span> <span class="o">==</span> <span class="s2">&quot;csr&quot;</span> <span class="ow">or</span> <span class="n">format</span> <span class="o">==</span> <span class="s2">&quot;csc&quot;</span>
            <span class="k">if</span> <span class="n">sparse</span><span class="p">:</span>
                <span class="n">psize</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">inputs</span><span class="p">[</span><span class="n">input_index</span><span class="p">]</span><span class="o">.</span><span class="n">nnz</span>
            <span class="bp">self</span><span class="o">.</span><span class="n">create_input_parameter</span><span class="p">(</span><span class="n">input_index</span><span class="p">,</span> <span class="n">psize</span><span class="p">,</span> <span class="n">dims</span><span class="p">,</span> <span class="n">sparse</span><span class="p">,</span> <span class="n">format</span><span class="p">)</span>
        <span class="bp">self</span><span class="o">.</span><span class="n">create_bias_parameter</span><span class="p">(</span><span class="n">bias</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">size</span><span class="p">)</span>
</pre></div>
</div>
<p>In network configuration, the layer can be specifies using the following code snippets. The arguments of this class are:</p>
<ul class="simple">
Y
Upload docs  
Yu Yang 已提交
432 433 434 435 436 
<li><code class="code docutils literal"><span class="pre">name</span></code> is the name identifier of the layer instance.</li>
<li><code class="code docutils literal"><span class="pre">type</span></code> is the type of the layer, specified using layer identifier.</li>
<li><code class="code docutils literal"><span class="pre">size</span></code> is the output size of the layer.</li>
<li><code class="code docutils literal"><span class="pre">bias</span></code> specifies whether this layer instance has bias.</li>
<li><code class="code docutils literal"><span class="pre">inputs</span></code> specifies a list of layer instance names as inputs.</li>
Y
Update docs  
Yu Yang 已提交
437 438 439 440 441 442 443 444 445 446 
</ul>
<div class="highlight-python"><div class="highlight"><pre><span></span><span class="n">Layer</span><span class="p">(</span>
    <span class="n">name</span> <span class="o">=</span> <span class="s2">&quot;fc1&quot;</span><span class="p">,</span>
    <span class="nb">type</span> <span class="o">=</span> <span class="s2">&quot;fc&quot;</span><span class="p">,</span>
    <span class="n">size</span> <span class="o">=</span> <span class="mi">64</span><span class="p">,</span>
    <span class="n">bias</span> <span class="o">=</span> <span class="bp">True</span><span class="p">,</span>
    <span class="n">inputs</span> <span class="o">=</span> <span class="p">[</span><span class="n">Input</span><span class="p">(</span><span class="s2">&quot;pool3&quot;</span><span class="p">)]</span>
<span class="p">)</span>
</pre></div>
</div>
Y
Upload docs  
Yu Yang 已提交
447 
<p>You are also recommended to implement a helper for the Python wrapper, which makes it easier to write models. You can refer to <code class="code docutils literal"><span class="pre">python/paddle/trainer_config_helpers/layers.py</span></code> for examples.</p>
Y
Update docs  
Yu Yang 已提交
448 449 450 451 452 453 454 455 456 457 458 459 460 
</div>


          </div>
        </div>
      </div>
      <div class="sphinxsidebar" role="navigation" aria-label="main navigation">
        <div class="sphinxsidebarwrapper">
  <h3><a href="../../index.html">Table Of Contents</a></h3>
  <ul>
<li><a class="reference internal" href="#">Writing New Layers</a></li>
<li><a class="reference internal" href="#derive-equations">Derive Equations</a></li>
<li><a class="reference internal" href="#implement-c-class">Implement C++ Class</a></li>
Y
Upload docs  
Yu Yang 已提交
461 
<li><a class="reference internal" href="#write-gradient-check-unit-test">Write Gradient Check Unit Test</a></li>
Y
Update docs  
Yu Yang 已提交
462 463 464 465 
<li><a class="reference internal" href="#implement-python-wrapper">Implement Python Wrapper</a></li>
</ul>

  <h4>Previous topic</h4>
Y
Upload docs  
Yu Yang 已提交
466 467 
  <p class="topless"><a href="index.html"
                        title="previous chapter">Writing New Layers</a></p>
Y
Update docs  
Yu Yang 已提交
468 
  <h4>Next topic</h4>
Y
Upload docs  
Yu Yang 已提交
469 470 
  <p class="topless"><a href="../../source/index.html"
                        title="next chapter">Source Code Documents</a></p>
Y
Update docs  
Yu Yang 已提交
471 472 473 
  <div role="note" aria-label="source link">
    <h3>This Page</h3>
    <ul class="this-page-menu">
Y
Upload docs  
Yu Yang 已提交
474 
      <li><a href="../../_sources/dev/new_layer/new_layer.txt"
Y
Update docs  
Yu Yang 已提交
475 476 477 478 479 480 
            rel="nofollow">Show Source</a></li>
    </ul>
   </div>
<div id="searchbox" style="display: none" role="search">
  <h3>Quick search</h3>
    <form class="search" action="../../search.html" method="get">
T
Deploy to GitHub Pages: 703cce35a871c78a5e3f1c33abf7fe788e2bbda5  
Travis CI 已提交
481 482 
      <div><input type="text" name="q" /></div>
      <div><input type="submit" value="Go" /></div>
Y
Update docs  
Yu Yang 已提交
483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 
      <input type="hidden" name="check_keywords" value="yes" />
      <input type="hidden" name="area" value="default" />
    </form>
</div>
<script type="text/javascript">$('#searchbox').show(0);</script>
        </div>
      </div>
      <div class="clearer"></div>
    </div>
    <div class="related" role="navigation" aria-label="related navigation">
      <h3>Navigation</h3>
      <ul>
        <li class="right" style="margin-right: 10px">
          <a href="../../genindex.html" title="General Index"
             >index</a></li>
        <li class="right" >
          <a href="../../py-modindex.html" title="Python Module Index"
             >modules</a> |</li>
        <li class="right" >
Y
Upload docs  
Yu Yang 已提交
502 
          <a href="../../source/index.html" title="Source Code Documents"
Y
Update docs  
Yu Yang 已提交
503 504 
             >next</a> |</li>
        <li class="right" >
Y
Upload docs  
Yu Yang 已提交
505 
          <a href="index.html" title="Writing New Layers"
Y
Update docs  
Yu Yang 已提交
506 
             >previous</a> |</li>
T
Deploy to GitHub Pages: 703cce35a871c78a5e3f1c33abf7fe788e2bbda5  
Travis CI 已提交
507 508 
        <li class="nav-item nav-item-0"><a href="../../index.html">PaddlePaddle  documentation</a> &#187;</li>
          <li class="nav-item nav-item-1"><a href="index.html" >Writing New Layers</a> &#187;</li>
Y
Update docs  
Yu Yang 已提交
509 510 511 
      </ul>
    </div>
    <div class="footer" role="contentinfo">
T
Deploy to GitHub Pages: 703cce35a871c78a5e3f1c33abf7fe788e2bbda5  
Travis CI 已提交
512 513 
        &#169; Copyright 2016, PaddlePaddle developers.
      Created using <a href="http://sphinx-doc.org/">Sphinx</a> 1.4.6.
Y
Update docs  
Yu Yang 已提交
514 515 516 
    </div>
  </body>
</html>