chore: import upstream snapshot with attribution
This commit is contained in:
wehub-resource-sync
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,382 @@
|
||||
<!DOCTYPE html>
|
||||
<html lang="zh">
|
||||
<head>
|
||||
<meta http-equiv="content-type" content="text/html;charset=utf-8"/>
|
||||
<meta name="viewport" content="width=device-width, initial-scale=1.0"/>
|
||||
<meta name="description" content="这是 PyTorch 中掩码语言模型的带注释的实现/教程。"/>
|
||||
|
||||
<meta name="twitter:card" content="summary"/>
|
||||
<meta name="twitter:image:src" content="https://avatars1.githubusercontent.com/u/64068543?s=400&v=4"/>
|
||||
<meta name="twitter:title" content="屏蔽语言模型"/>
|
||||
<meta name="twitter:description" content="这是 PyTorch 中掩码语言模型的带注释的实现/教程。"/>
|
||||
<meta name="twitter:site" content="@labmlai"/>
|
||||
<meta name="twitter:creator" content="@labmlai"/>
|
||||
|
||||
<meta property="og:url" content="https://nn.labml.ai/transformers/mlm/index.html"/>
|
||||
<meta property="og:title" content="屏蔽语言模型"/>
|
||||
<meta property="og:image" content="https://avatars1.githubusercontent.com/u/64068543?s=400&v=4"/>
|
||||
<meta property="og:site_name" content="屏蔽语言模型"/>
|
||||
<meta property="og:type" content="object"/>
|
||||
<meta property="og:title" content="屏蔽语言模型"/>
|
||||
<meta property="og:description" content="这是 PyTorch 中掩码语言模型的带注释的实现/教程。"/>
|
||||
|
||||
<title>屏蔽语言模型</title>
|
||||
<link rel="shortcut icon" href="/icon.png"/>
|
||||
<link rel="stylesheet" href="../../pylit.css?v=1">
|
||||
<link rel="canonical" href="https://nn.labml.ai/transformers/mlm/index.html"/>
|
||||
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.13.18/dist/katex.min.css" integrity="sha384-zTROYFVGOfTw7JV7KUu8udsvW2fx4lWOsCEDqhBreBwlHI4ioVRtmIvEThzJHGET" crossorigin="anonymous">
|
||||
|
||||
<!-- Global site tag (gtag.js) - Google Analytics -->
|
||||
<script async src="https://www.googletagmanager.com/gtag/js?id=G-4V3HC8HBLH"></script>
|
||||
<script>
|
||||
window.dataLayer = window.dataLayer || [];
|
||||
|
||||
function gtag() {
|
||||
dataLayer.push(arguments);
|
||||
}
|
||||
|
||||
gtag('js', new Date());
|
||||
|
||||
gtag('config', 'G-4V3HC8HBLH');
|
||||
</script>
|
||||
</head>
|
||||
<body>
|
||||
<div id='container'>
|
||||
<div id="background"></div>
|
||||
<div class='section'>
|
||||
<div class='docs'>
|
||||
<p>
|
||||
<a class="parent" href="/">home</a>
|
||||
<a class="parent" href="../index.html">transformers</a>
|
||||
<a class="parent" href="index.html">mlm</a>
|
||||
</p>
|
||||
<p>
|
||||
<a href="https://github.com/labmlai/annotated_deep_learning_paper_implementations" target="_blank">
|
||||
<img alt="Github"
|
||||
src="https://img.shields.io/github/stars/labmlai/annotated_deep_learning_paper_implementations?style=social"
|
||||
style="max-width:100%;"/></a>
|
||||
<a href="https://twitter.com/labmlai" rel="nofollow" target="_blank">
|
||||
<img alt="Twitter"
|
||||
src="https://img.shields.io/twitter/follow/labmlai?style=social"
|
||||
style="max-width:100%;"/></a>
|
||||
</p>
|
||||
<p>
|
||||
<a href="https://github.com/labmlai/annotated_deep_learning_paper_implementations/tree/master/labml_nn/transformers/mlm/__init__.py" target="_blank">
|
||||
View code on Github</a>
|
||||
</p>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-0'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-0'>#</a>
|
||||
</div>
|
||||
<h1>蒙面语言模型 (MLM)</h1>
|
||||
<p>这是 <a href="https://pytorch.org">PyTorch</a> 实现的掩码语言模型 (MLM)，用于预训白文《BERT<a href="https://arxiv.org/abs/1810.04805">：深度双向转换器预训练 BERT》中介绍的</a> BERT 模型。</p>
|
||||
<h2>BERT 预训练</h2>
|
||||
<p>BERT 模型是变压器模型。本文使用 MLM 和下一句预测对模型进行了预训练。我们只在这里实施了传销。</p>
|
||||
<h3>下一句预测</h3>
|
||||
<p>在<em>下一个句子预测</em>中，给出两个句子，<code class="highlight"><span></span><span class="n">A</span></code>
|
||||
<code class="highlight"><span></span><span class="n">B</span></code>
|
||||
然后模型对实际文本<code class="highlight"><span></span><span class="n">A</span></code>
|
||||
中后面的句子是否<code class="highlight"><span></span><span class="n">B</span></code>
|
||||
是后面的句子进行二进制预测。该模型有 50% 的时间为实际句子对，50% 的时间为随机句对。这种分类是在应用传销时完成的。<em>我们还没有在这里实现这一点。</em></p>
|
||||
<h2>Masked LM</h2>
|
||||
<p>这会随机掩盖一定比例的代币，并训练模型预测被掩码的代币。他们通过用特殊<strong>代币替换15％的代<code class="highlight"><span></span><span class="p">[</span><span class="n">MASK</span><span class="p">]</span></code>
|
||||
币来掩盖</strong>它们。</p>
|
||||
<p>损失仅通过预测被掩码的代币来计算。这在微调和实际使用过程中会导致问题，因为当时没有<code class="highlight"><span></span><span class="p">[</span><span class="n">MASK</span><span class="p">]</span></code>
|
||||
令牌。因此，我们可能得不到任何有意义的陈述。</p>
|
||||
<p>为了克服这个问题<strong>，10％的蒙面代币被替换为原始代币</strong>，另外 <strong>10％的蒙面代币被随机代币所取代</strong>。无论该位置的输入代币是否为，这都会训练模型给出有关实际代币的表现形式<code class="highlight"><span></span><span class="p">[</span><span class="n">MASK</span><span class="p">]</span></code>
|
||||
。用随机代币替换会使它给出的表现形式也包含来自上下文的信息；因为它必须使用上下文来修复随机替换的标记。</p>
|
||||
<h2>训练</h2>
|
||||
<p>MLM 比自回归模型更难训练，因为它们的训练信号较小。也就是说，每个样本只训练了一小部分的预测。</p>
|
||||
<p>另一个问题是，由于该模型是双向的，因此任何代币都可以看到任何其他代币。这使得 “信用分配” 变得更加困难。假设你有角色等级模型想要预测<code class="highlight"><span></span><span class="n">home</span> <span class="o">*</span><span class="n">s</span> <span class="n">where</span> <span class="n">i</span> <span class="n">want</span> <span class="n">to</span> <span class="n">be</span></code>
|
||||
。至少在训练的早期阶段，很难弄清楚为什么要用<code class="highlight"><span></span><span class="o">*</span></code>
|
||||
它来代替<code class="highlight"><span></span><span class="n">i</span></code>
|
||||
，可能是整句话中的任何东西。而在自回归环境中，模型只<code class="highlight"><span></span><span class="n">h</span></code>
|
||||
需要用于预测<code class="highlight"><span></span><span class="n">o</span></code>
|
||||
<code class="highlight"><span></span><span class="n">e</span></code>
|
||||
和<code class="highlight"><span></span><span class="n">hom</span></code>
|
||||
预测等等。因此，该模型最初将首先在较短的上下文中开始预测，然后学会使用较长的上下文进行预测。由于 MLM 有这个问题，如果你一开始使用较小的序列长度，然后再使用更长的序列长度，那么训练速度会快得多。</p>
|
||||
<p>这是简单 MLM 模型的<a href="experiment.html">训练代码</a>。</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">65</span><span></span><span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">List</span>
|
||||
<span class="lineno">66</span>
|
||||
<span class="lineno">67</span><span class="kn">import</span> <span class="nn">torch</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-1'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-1'>#</a>
|
||||
</div>
|
||||
<h2>Masked LM (传销)</h2>
|
||||
<p>该类实现给定批次令牌序列的掩码过程。</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">70</span><span class="k">class</span> <span class="nc">MLM</span><span class="p">:</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-2'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-2'>#</a>
|
||||
</div>
|
||||
<ul><li><code class="highlight"><span></span><span class="n">padding_token</span></code>
|
||||
是填充令牌<code class="highlight"><span></span><span class="p">[</span><span class="n">PAD</span><span class="p">]</span></code>
|
||||
。我们将用它来标记不应用于损失计算的标签。</li>
|
||||
<li><code class="highlight"><span></span><span class="n">mask_token</span></code>
|
||||
是掩码令牌<code class="highlight"><span></span><span class="p">[</span><span class="n">MASK</span><span class="p">]</span></code>
|
||||
。</li>
|
||||
<li><code class="highlight"><span></span><span class="n">no_mask_tokens</span></code>
|
||||
是不应该被屏蔽的令牌的列表。如果我们同时使用分类等其他任务来训练 MLM，并且我们有这样的令牌不应该被掩盖，那么<code class="highlight"><span></span><span class="p">[</span><span class="n">CLS</span><span class="p">]</span></code>
|
||||
这很有用。</li>
|
||||
<li><code class="highlight"><span></span><span class="n">n_tokens</span></code>
|
||||
代币总数（用于生成随机令牌）</li>
|
||||
<li><code class="highlight"><span></span><span class="n">masking_prob</span></code>
|
||||
是掩蔽概率</li>
|
||||
<li><code class="highlight"><span></span><span class="n">randomize_prob</span></code>
|
||||
是用随机代币替换的概率</li>
|
||||
<li><code class="highlight"><span></span><span class="n">no_change_prob</span></code>
|
||||
是用原始代币替换的概率</li></ul>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">77</span> <span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="o">*</span><span class="p">,</span>
|
||||
<span class="lineno">78</span> <span class="n">padding_token</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span> <span class="n">mask_token</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span> <span class="n">no_mask_tokens</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="nb">int</span><span class="p">],</span> <span class="n">n_tokens</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
|
||||
<span class="lineno">79</span> <span class="n">masking_prob</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">0.15</span><span class="p">,</span> <span class="n">randomize_prob</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">0.1</span><span class="p">,</span> <span class="n">no_change_prob</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">0.1</span><span class="p">,</span>
|
||||
<span class="lineno">80</span> <span class="p">):</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-3'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-3'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">93</span> <span class="bp">self</span><span class="o">.</span><span class="n">n_tokens</span> <span class="o">=</span> <span class="n">n_tokens</span>
|
||||
<span class="lineno">94</span> <span class="bp">self</span><span class="o">.</span><span class="n">no_change_prob</span> <span class="o">=</span> <span class="n">no_change_prob</span>
|
||||
<span class="lineno">95</span> <span class="bp">self</span><span class="o">.</span><span class="n">randomize_prob</span> <span class="o">=</span> <span class="n">randomize_prob</span>
|
||||
<span class="lineno">96</span> <span class="bp">self</span><span class="o">.</span><span class="n">masking_prob</span> <span class="o">=</span> <span class="n">masking_prob</span>
|
||||
<span class="lineno">97</span> <span class="bp">self</span><span class="o">.</span><span class="n">no_mask_tokens</span> <span class="o">=</span> <span class="n">no_mask_tokens</span> <span class="o">+</span> <span class="p">[</span><span class="n">padding_token</span><span class="p">,</span> <span class="n">mask_token</span><span class="p">]</span>
|
||||
<span class="lineno">98</span> <span class="bp">self</span><span class="o">.</span><span class="n">padding_token</span> <span class="o">=</span> <span class="n">padding_token</span>
|
||||
<span class="lineno">99</span> <span class="bp">self</span><span class="o">.</span><span class="n">mask_token</span> <span class="o">=</span> <span class="n">mask_token</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-4'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-4'>#</a>
|
||||
</div>
|
||||
<ul><li><code class="highlight"><span></span><span class="n">x</span></code>
|
||||
是一批输入令牌序列。它是一个<code class="highlight"><span></span><span class="n">long</span></code>
|
||||
带形状的张量<code class="highlight"><span></span><span class="p">[</span><span class="n">seq_len</span><span class="p">,</span> <span class="n">batch_size</span><span class="p">]</span></code>
|
||||
。</li></ul>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">101</span> <span class="k">def</span> <span class="fm">__call__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">):</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-5'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-5'>#</a>
|
||||
</div>
|
||||
<p>代币<code class="highlight"><span></span><span class="n">masking_prob</span></code>
|
||||
面具</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">108</span> <span class="n">full_mask</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="n">x</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">x</span><span class="o">.</span><span class="n">device</span><span class="p">)</span> <span class="o"><</span> <span class="bp">self</span><span class="o">.</span><span class="n">masking_prob</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-6'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-6'>#</a>
|
||||
</div>
|
||||
<p>揭露面具<code class="highlight"><span></span><span class="n">no_mask_tokens</span></code>
|
||||
</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">110</span> <span class="k">for</span> <span class="n">t</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">no_mask_tokens</span><span class="p">:</span>
|
||||
<span class="lineno">111</span> <span class="n">full_mask</span> <span class="o">&=</span> <span class="n">x</span> <span class="o">!=</span> <span class="n">t</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-7'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-7'>#</a>
|
||||
</div>
|
||||
<p>将令牌替换为原始令牌的掩码</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">114</span> <span class="n">unchanged</span> <span class="o">=</span> <span class="n">full_mask</span> <span class="o">&</span> <span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="n">x</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">x</span><span class="o">.</span><span class="n">device</span><span class="p">)</span> <span class="o"><</span> <span class="bp">self</span><span class="o">.</span><span class="n">no_change_prob</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-8'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-8'>#</a>
|
||||
</div>
|
||||
<p>将令牌替换为随机令牌的掩码</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">116</span> <span class="n">random_token_mask</span> <span class="o">=</span> <span class="n">full_mask</span> <span class="o">&</span> <span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="n">x</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">x</span><span class="o">.</span><span class="n">device</span><span class="p">)</span> <span class="o"><</span> <span class="bp">self</span><span class="o">.</span><span class="n">randomize_prob</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-9'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-9'>#</a>
|
||||
</div>
|
||||
<p>要替换为随机令牌的令牌索引</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">118</span> <span class="n">random_token_idx</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nonzero</span><span class="p">(</span><span class="n">random_token_mask</span><span class="p">,</span> <span class="n">as_tuple</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-10'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-10'>#</a>
|
||||
</div>
|
||||
<p>每个地点的随机代币</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">120</span> <span class="n">random_tokens</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randint</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">n_tokens</span><span class="p">,</span> <span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="n">random_token_idx</span><span class="p">[</span><span class="mi">0</span><span class="p">]),),</span> <span class="n">device</span><span class="o">=</span><span class="n">x</span><span class="o">.</span><span class="n">device</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-11'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-11'>#</a>
|
||||
</div>
|
||||
<p>最后一组将被替换的令牌<code class="highlight"><span></span><span class="p">[</span><span class="n">MASK</span><span class="p">]</span></code>
|
||||
</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">122</span> <span class="n">mask</span> <span class="o">=</span> <span class="n">full_mask</span> <span class="o">&</span> <span class="o">~</span><span class="n">random_token_mask</span> <span class="o">&</span> <span class="o">~</span><span class="n">unchanged</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-12'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-12'>#</a>
|
||||
</div>
|
||||
<p>克隆标签的输入</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">125</span> <span class="n">y</span> <span class="o">=</span> <span class="n">x</span><span class="o">.</span><span class="n">clone</span><span class="p">()</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-13'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-13'>#</a>
|
||||
</div>
|
||||
<p>替换为<code class="highlight"><span></span><span class="p">[</span><span class="n">MASK</span><span class="p">]</span></code>
|
||||
令牌；请注意，这不包括原始令牌保持不变的令牌和被随机令牌替换的令牌。</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">130</span> <span class="n">x</span><span class="o">.</span><span class="n">masked_fill_</span><span class="p">(</span><span class="n">mask</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">mask_token</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-14'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-14'>#</a>
|
||||
</div>
|
||||
<p>分配随机代币</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">132</span> <span class="n">x</span><span class="p">[</span><span class="n">random_token_idx</span><span class="p">]</span> <span class="o">=</span> <span class="n">random_tokens</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-15'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-15'>#</a>
|
||||
</div>
|
||||
<p>将令牌<code class="highlight"><span></span><span class="p">[</span><span class="n">PAD</span><span class="p">]</span></code>
|
||||
分配给标签中的所有其他位置。等于的标签<code class="highlight"><span></span><span class="p">[</span><span class="n">PAD</span><span class="p">]</span></code>
|
||||
将不会用于亏损。</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">136</span> <span class="n">y</span><span class="o">.</span><span class="n">masked_fill_</span><span class="p">(</span><span class="o">~</span><span class="n">full_mask</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">padding_token</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-16'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-16'>#</a>
|
||||
</div>
|
||||
<p>返回屏蔽的输入和标签</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">139</span> <span class="k">return</span> <span class="n">x</span><span class="p">,</span> <span class="n">y</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='footer'>
|
||||
<a href="https://labml.ai">labml.ai</a>
|
||||
</div>
|
||||
</div>
|
||||
<script src=../../interactive.js?v=1"></script>
|
||||
<script>
|
||||
function handleImages() {
|
||||
var images = document.querySelectorAll('p>img')
|
||||
|
||||
for (var i = 0; i < images.length; ++i) {
|
||||
handleImage(images[i])
|
||||
}
|
||||
}
|
||||
|
||||
function handleImage(img) {
|
||||
img.parentElement.style.textAlign = 'center'
|
||||
|
||||
var modal = document.createElement('div')
|
||||
modal.id = 'modal'
|
||||
|
||||
var modalContent = document.createElement('div')
|
||||
modal.appendChild(modalContent)
|
||||
|
||||
var modalImage = document.createElement('img')
|
||||
modalContent.appendChild(modalImage)
|
||||
|
||||
var span = document.createElement('span')
|
||||
span.classList.add('close')
|
||||
span.textContent = 'x'
|
||||
modal.appendChild(span)
|
||||
|
||||
img.onclick = function () {
|
||||
console.log('clicked')
|
||||
document.body.appendChild(modal)
|
||||
modalImage.src = img.src
|
||||
}
|
||||
|
||||
span.onclick = function () {
|
||||
document.body.removeChild(modal)
|
||||
}
|
||||
}
|
||||
|
||||
handleImages()
|
||||
</script>
|
||||
</body>
|
||||
</html>
|
||||
@@ -0,0 +1,153 @@
|
||||
<!DOCTYPE html>
|
||||
<html lang="zh">
|
||||
<head>
|
||||
<meta http-equiv="content-type" content="text/html;charset=utf-8"/>
|
||||
<meta name="viewport" content="width=device-width, initial-scale=1.0"/>
|
||||
<meta name="description" content=""/>
|
||||
|
||||
<meta name="twitter:card" content="summary"/>
|
||||
<meta name="twitter:image:src" content="https://avatars1.githubusercontent.com/u/64068543?s=400&v=4"/>
|
||||
<meta name="twitter:title" content="屏蔽语言模型 (MLM)"/>
|
||||
<meta name="twitter:description" content=""/>
|
||||
<meta name="twitter:site" content="@labmlai"/>
|
||||
<meta name="twitter:creator" content="@labmlai"/>
|
||||
|
||||
<meta property="og:url" content="https://nn.labml.ai/transformers/mlm/readme.html"/>
|
||||
<meta property="og:title" content="屏蔽语言模型 (MLM)"/>
|
||||
<meta property="og:image" content="https://avatars1.githubusercontent.com/u/64068543?s=400&v=4"/>
|
||||
<meta property="og:site_name" content="屏蔽语言模型 (MLM)"/>
|
||||
<meta property="og:type" content="object"/>
|
||||
<meta property="og:title" content="屏蔽语言模型 (MLM)"/>
|
||||
<meta property="og:description" content=""/>
|
||||
|
||||
<title>屏蔽语言模型 (MLM)</title>
|
||||
<link rel="shortcut icon" href="/icon.png"/>
|
||||
<link rel="stylesheet" href="../../pylit.css?v=1">
|
||||
<link rel="canonical" href="https://nn.labml.ai/transformers/mlm/readme.html"/>
|
||||
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.13.18/dist/katex.min.css" integrity="sha384-zTROYFVGOfTw7JV7KUu8udsvW2fx4lWOsCEDqhBreBwlHI4ioVRtmIvEThzJHGET" crossorigin="anonymous">
|
||||
|
||||
<!-- Global site tag (gtag.js) - Google Analytics -->
|
||||
<script async src="https://www.googletagmanager.com/gtag/js?id=G-4V3HC8HBLH"></script>
|
||||
<script>
|
||||
window.dataLayer = window.dataLayer || [];
|
||||
|
||||
function gtag() {
|
||||
dataLayer.push(arguments);
|
||||
}
|
||||
|
||||
gtag('js', new Date());
|
||||
|
||||
gtag('config', 'G-4V3HC8HBLH');
|
||||
</script>
|
||||
</head>
|
||||
<body>
|
||||
<div id='container'>
|
||||
<div id="background"></div>
|
||||
<div class='section'>
|
||||
<div class='docs'>
|
||||
<p>
|
||||
<a class="parent" href="/">home</a>
|
||||
<a class="parent" href="../index.html">transformers</a>
|
||||
<a class="parent" href="index.html">mlm</a>
|
||||
</p>
|
||||
<p>
|
||||
<a href="https://github.com/labmlai/annotated_deep_learning_paper_implementations" target="_blank">
|
||||
<img alt="Github"
|
||||
src="https://img.shields.io/github/stars/labmlai/annotated_deep_learning_paper_implementations?style=social"
|
||||
style="max-width:100%;"/></a>
|
||||
<a href="https://twitter.com/labmlai" rel="nofollow" target="_blank">
|
||||
<img alt="Twitter"
|
||||
src="https://img.shields.io/twitter/follow/labmlai?style=social"
|
||||
style="max-width:100%;"/></a>
|
||||
</p>
|
||||
<p>
|
||||
<a href="https://github.com/labmlai/annotated_deep_learning_paper_implementations/tree/master/labml_nn/transformers/mlm/readme.md" target="_blank">
|
||||
View code on Github</a>
|
||||
</p>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-0'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-0'>#</a>
|
||||
</div>
|
||||
<h1><a href="https://nn.labml.ai/transformers/mlm/index.html">蒙面语言模型 (MLM)</a></h1>
|
||||
<p>这是掩码语言模型 (MLM) 的 <a href="https://pytorch.org">PyTorch</a> 实现，用于预训白文《BERT<a href="https://arxiv.org/abs/1810.04805">：预训练深度双向转换器以促进语言理解》中介绍的 BER</a> T 模型。</p>
|
||||
<h2>BERT 预训练</h2>
|
||||
<p>BERT 模型是变压器模型。本文使用 MLM 和下一句预测对模型进行了预训练。我们只在这里实施了传销。</p>
|
||||
<h3>下一句预测</h3>
|
||||
<p>在<em>下一个句子预测</em>中，给出两个句子，<code class="highlight"><span></span><span class="n">A</span></code>
|
||||
<code class="highlight"><span></span><span class="n">B</span></code>
|
||||
然后模型对实际文本<code class="highlight"><span></span><span class="n">A</span></code>
|
||||
中后面的句子是否<code class="highlight"><span></span><span class="n">B</span></code>
|
||||
是后面的句子进行二进制预测。该模型有 50% 的时间为实际句子对，50% 的时间为随机句对。这种分类是在应用传销时完成的。<em>我们还没有在这里实现这一点。</em></p>
|
||||
<h2>Masked LM</h2>
|
||||
<p>这会随机掩盖一定比例的代币，并训练模型预测被掩码的代币。他们通过用特殊<strong>代币替换15％的代<code class="highlight"><span></span><span class="p">[</span><span class="n">MASK</span><span class="p">]</span></code>
|
||||
币来掩盖</strong>它们。</p>
|
||||
<p>损失仅通过预测被掩码的代币来计算。这在微调和实际使用过程中会导致问题，因为当时没有<code class="highlight"><span></span><span class="p">[</span><span class="n">MASK</span><span class="p">]</span></code>
|
||||
令牌。因此，我们可能得不到任何有意义的陈述。</p>
|
||||
<p>为了克服这个问题<strong>，10％的蒙面代币被替换为原始代币</strong>，另外 <strong>10％的蒙面代币被随机代币所取代</strong>。无论该位置的输入代币是否为，这都会训练模型给出有关实际代币的表现形式<code class="highlight"><span></span><span class="p">[</span><span class="n">MASK</span><span class="p">]</span></code>
|
||||
。用随机代币替换会使它给出的表现形式也包含来自上下文的信息；因为它必须使用上下文来修复随机替换的标记。</p>
|
||||
<h2>训练</h2>
|
||||
<p>MLM 比自回归模型更难训练，因为它们的训练信号较小。也就是说，每个样本只训练了一小部分的预测。</p>
|
||||
<p>另一个问题是，由于该模型是双向的，因此任何代币都可以看到任何其他代币。这使得 “信用分配” 变得更加困难。假设你有角色等级模型想要预测<code class="highlight"><span></span><span class="n">home</span> <span class="o">*</span><span class="n">s</span> <span class="n">where</span> <span class="n">i</span> <span class="n">want</span> <span class="n">to</span> <span class="n">be</span></code>
|
||||
。至少在训练的早期阶段，很难弄清楚为什么要用<code class="highlight"><span></span><span class="o">*</span></code>
|
||||
它来代替<code class="highlight"><span></span><span class="n">i</span></code>
|
||||
，可能是整句话中的任何东西。而在自回归环境中，模型只<code class="highlight"><span></span><span class="n">h</span></code>
|
||||
需要用于预测<code class="highlight"><span></span><span class="n">o</span></code>
|
||||
<code class="highlight"><span></span><span class="n">e</span></code>
|
||||
和<code class="highlight"><span></span><span class="n">hom</span></code>
|
||||
预测等等。因此，该模型最初将首先在较短的上下文中开始预测，然后学会使用较长的上下文进行预测。由于 MLM 有这个问题，如果你一开始使用较小的序列长度，然后再使用更长的序列长度，那么训练速度会快得多。</p>
|
||||
<p>这是简单 MLM 模型的<a href="https://nn.labml.ai/transformers/mlm/experiment.html">训练代码</a>。</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
|
||||
</div>
|
||||
</div>
|
||||
<div class='footer'>
|
||||
<a href="https://labml.ai">labml.ai</a>
|
||||
</div>
|
||||
</div>
|
||||
<script src=../../interactive.js?v=1"></script>
|
||||
<script>
|
||||
function handleImages() {
|
||||
var images = document.querySelectorAll('p>img')
|
||||
|
||||
for (var i = 0; i < images.length; ++i) {
|
||||
handleImage(images[i])
|
||||
}
|
||||
}
|
||||
|
||||
function handleImage(img) {
|
||||
img.parentElement.style.textAlign = 'center'
|
||||
|
||||
var modal = document.createElement('div')
|
||||
modal.id = 'modal'
|
||||
|
||||
var modalContent = document.createElement('div')
|
||||
modal.appendChild(modalContent)
|
||||
|
||||
var modalImage = document.createElement('img')
|
||||
modalContent.appendChild(modalImage)
|
||||
|
||||
var span = document.createElement('span')
|
||||
span.classList.add('close')
|
||||
span.textContent = 'x'
|
||||
modal.appendChild(span)
|
||||
|
||||
img.onclick = function () {
|
||||
console.log('clicked')
|
||||
document.body.appendChild(modal)
|
||||
modalImage.src = img.src
|
||||
}
|
||||
|
||||
span.onclick = function () {
|
||||
document.body.removeChild(modal)
|
||||
}
|
||||
}
|
||||
|
||||
handleImages()
|
||||
</script>
|
||||
</body>
|
||||
</html>
|
||||
Reference in New Issue
Block a user