mirror of
https://github.com/labmlai/annotated_deep_learning_paper_implementations.git
synced 2025-08-06 15:22:21 +08:00
751 lines
50 KiB
HTML
751 lines
50 KiB
HTML
<!DOCTYPE html>
|
|
<html>
|
|
<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="This is an annotated implementation/tutorial a miniature version of Switch Transformer in PyTorch."/>
|
|
|
|
<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="Switch Transformer"/>
|
|
<meta name="twitter:description" content="This is an annotated implementation/tutorial a miniature version of Switch Transformer in PyTorch."/>
|
|
<meta name="twitter:site" content="@labmlai"/>
|
|
<meta name="twitter:creator" content="@labmlai"/>
|
|
|
|
<meta property="og:url" content="https://nn.labml.ai/transformers/switch/index.html"/>
|
|
<meta property="og:title" content="Switch Transformer"/>
|
|
<meta property="og:image" content="https://avatars1.githubusercontent.com/u/64068543?s=400&v=4"/>
|
|
<meta property="og:site_name" content="LabML Neural Networks"/>
|
|
<meta property="og:type" content="object"/>
|
|
<meta property="og:title" content="Switch Transformer"/>
|
|
<meta property="og:description" content="This is an annotated implementation/tutorial a miniature version of Switch Transformer in PyTorch."/>
|
|
|
|
<title>Switch Transformer</title>
|
|
<link rel="shortcut icon" href="/icon.png"/>
|
|
<link rel="stylesheet" href="../../pylit.css">
|
|
<link rel="canonical" href="https://nn.labml.ai/transformers/switch/index.html"/>
|
|
<!-- 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">switch</a>
|
|
</p>
|
|
<p>
|
|
|
|
<a href="https://github.com/lab-ml/labml_nn/tree/master/labml_nn/transformers/switch/__init__.py">
|
|
<img alt="Github"
|
|
src="https://img.shields.io/github/stars/lab-ml/nn?style=social"
|
|
style="max-width:100%;"/></a>
|
|
<a href="https://twitter.com/labmlai"
|
|
rel="nofollow">
|
|
<img alt="Twitter"
|
|
src="https://img.shields.io/twitter/follow/labmlai?style=social"
|
|
style="max-width:100%;"/></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>Switch Transformer</h1>
|
|
<p>This is a miniature <a href="https://pytorch.org">PyTorch</a> implementation of the paper
|
|
<a href="https://arxiv.org/abs/2101.03961">Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity</a>.
|
|
Our implementation only has a few million parameters and doesn’t do model parallel distributed training.
|
|
It does single GPU training, but we implement the concept of switching as described in the paper.</p>
|
|
<p>The Switch Transformer uses different parameters for each token by switching among parameters
|
|
based on the token.
|
|
Therefore, only a fraction of parameters are chosen for each token.
|
|
So you can have more parameters but less computational cost.</p>
|
|
<p>The switching happens at the Position-wise Feedforward network (FFN) of each transformer block.
|
|
Position-wise feedforward network consists of two sequentially fully connected layers.
|
|
In switch transformer we have multiple FFNs (multiple experts),
|
|
and we chose which one to use based on a router.
|
|
The output is a set of probabilities for picking a FFN,
|
|
and we pick the one with the highest probability and only evaluate that.
|
|
So essentially the computational cost is the same as having a single FFN.
|
|
In our implementation this doesn’t parallelize well when you have many or large FFNs since it’s all
|
|
happening on a single GPU.
|
|
In a distributed setup you would have each FFN (each very large) on a different device.</p>
|
|
<p>The paper introduces another loss term to balance load among the experts (FFNs) and
|
|
discusses dropping tokens when routing is not balanced.</p>
|
|
<p>Here’s <a href="experiment.html">the training code</a> and a notebook for training a switch transformer on Tiny Shakespeare dataset.</p>
|
|
<p><a href="https://colab.research.google.com/github/lab-ml/nn/blob/master/labml_nn/transformers/switch/experiment.ipynb"><img alt="Open In Colab" src="https://colab.research.google.com/assets/colab-badge.svg" /></a>
|
|
<a href="https://app.labml.ai/run/c4656c605b9311eba13d0242ac1c0002"><img alt="View Run" src="https://img.shields.io/badge/labml-experiment-brightgreen" /></a></p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">40</span><span></span><span class="kn">import</span> <span class="nn">torch</span>
|
|
<span class="lineno">41</span><span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">nn</span>
|
|
<span class="lineno">42</span>
|
|
<span class="lineno">43</span><span class="kn">from</span> <span class="nn">labml_helpers.module</span> <span class="kn">import</span> <span class="n">Module</span>
|
|
<span class="lineno">44</span><span class="kn">from</span> <span class="nn">labml_nn.transformers.mha</span> <span class="kn">import</span> <span class="n">MultiHeadAttention</span>
|
|
<span class="lineno">45</span><span class="kn">from</span> <span class="nn">labml_nn.transformers.feed_forward</span> <span class="kn">import</span> <span class="n">FeedForward</span>
|
|
<span class="lineno">46</span><span class="kn">from</span> <span class="nn">labml_nn.utils</span> <span class="kn">import</span> <span class="n">clone_module_list</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>Routing among multiple FFNs</h2>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">49</span><span class="k">class</span> <span class="nc">SwitchFeedForward</span><span class="p">(</span><span class="n">Module</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>capacity_factor</code> is the capacity of each expert as a factor relative to ideally balanced load</li>
|
|
<li><code>drop_tokens</code> specifies whether to drop tokens if more tokens are routed to an expert than the capacity</li>
|
|
<li><code>is_scale_prob</code> specifies whether to multiply the input to the FFN by the routing probability</li>
|
|
<li><code>n_experts</code> is the number of experts</li>
|
|
<li><code>expert</code> is the expert layer, a <a href="../feed_forward.html">FFN module</a></li>
|
|
<li><code>d_model</code> is the number of features in a token embedding</li>
|
|
<li><code>d_ff</code> is the number of features in the hidden layer of the FFN</li>
|
|
<li><code>dropout</code> is dropout probability in the FFN</li>
|
|
</ul>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">54</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">55</span> <span class="n">capacity_factor</span><span class="p">:</span> <span class="nb">float</span><span class="p">,</span>
|
|
<span class="lineno">56</span> <span class="n">drop_tokens</span><span class="p">:</span> <span class="nb">bool</span><span class="p">,</span>
|
|
<span class="lineno">57</span> <span class="n">is_scale_prob</span><span class="p">:</span> <span class="nb">bool</span><span class="p">,</span>
|
|
<span class="lineno">58</span> <span class="n">n_experts</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
|
|
<span class="lineno">59</span> <span class="n">expert</span><span class="p">:</span> <span class="n">FeedForward</span><span class="p">,</span>
|
|
<span class="lineno">60</span> <span class="n">d_model</span><span class="p">:</span> <span class="nb">int</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">71</span> <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
|
|
<span class="lineno">72</span>
|
|
<span class="lineno">73</span> <span class="bp">self</span><span class="o">.</span><span class="n">capacity_factor</span> <span class="o">=</span> <span class="n">capacity_factor</span>
|
|
<span class="lineno">74</span> <span class="bp">self</span><span class="o">.</span><span class="n">is_scale_prob</span> <span class="o">=</span> <span class="n">is_scale_prob</span>
|
|
<span class="lineno">75</span> <span class="bp">self</span><span class="o">.</span><span class="n">n_experts</span> <span class="o">=</span> <span class="n">n_experts</span>
|
|
<span class="lineno">76</span> <span class="bp">self</span><span class="o">.</span><span class="n">drop_tokens</span> <span class="o">=</span> <span class="n">drop_tokens</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-4'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-4'>#</a>
|
|
</div>
|
|
<p>make copies of the FFNs</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">79</span> <span class="bp">self</span><span class="o">.</span><span class="n">experts</span> <span class="o">=</span> <span class="n">clone_module_list</span><span class="p">(</span><span class="n">expert</span><span class="p">,</span> <span class="n">n_experts</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>Routing layer and softmax</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">81</span> <span class="bp">self</span><span class="o">.</span><span class="n">switch</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="n">d_model</span><span class="p">,</span> <span class="n">n_experts</span><span class="p">)</span>
|
|
<span class="lineno">82</span> <span class="bp">self</span><span class="o">.</span><span class="n">softmax</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Softmax</span><span class="p">(</span><span class="n">dim</span><span class="o">=-</span><span class="mi">1</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-6'>
|
|
<div class='docs doc-strings'>
|
|
<div class='section-link'>
|
|
<a href='#section-6'>#</a>
|
|
</div>
|
|
<ul>
|
|
<li><code>x</code> is the input to the switching module with shape <code>[seq_len, batch_size, d_model]</code></li>
|
|
</ul>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">84</span> <span class="k">def</span> <span class="nf">forward</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-7'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-7'>#</a>
|
|
</div>
|
|
<p>Capture the shape to change shapes later</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">90</span> <span class="n">seq_len</span><span class="p">,</span> <span class="n">batch_size</span><span class="p">,</span> <span class="n">d_model</span> <span class="o">=</span> <span class="n">x</span><span class="o">.</span><span class="n">shape</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>Flatten the sequence and batch dimensions</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">92</span> <span class="n">x</span> <span class="o">=</span> <span class="n">x</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="n">d_model</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>Get routing probabilities for each of the tokens.
|
|
<script type="math/tex; mode=display">p_i(x) = \frac{e^{h(x)_i}}{\sum^N_j e^{h(x)_j}}</script>
|
|
where $N$ is the number of experts <code>n_experts</code> and
|
|
$h(\cdot)$ is the linear transformation of token embeddings.</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">98</span> <span class="n">route_prob</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">softmax</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">switch</span><span class="p">(</span><span class="n">x</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>Get the maximum routing probabilities and the routes.
|
|
We route to the expert with highest probability</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">102</span> <span class="n">route_prob_max</span><span class="p">,</span> <span class="n">routes</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="n">route_prob</span><span class="p">,</span> <span class="n">dim</span><span class="o">=-</span><span class="mi">1</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>Scale the inputs to the experts by the routing probabilities</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">105</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">is_scale_prob</span><span class="p">:</span>
|
|
<span class="lineno">106</span> <span class="n">factor</span> <span class="o">=</span> <span class="n">route_prob_max</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>Don’t scale the values but multiply by $\frac{p}{\hat{p}} = 1$ so that the gradients flow</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">108</span> <span class="k">else</span><span class="p">:</span>
|
|
<span class="lineno">109</span> <span class="n">factor</span> <span class="o">=</span> <span class="n">route_prob_max</span> <span class="o">/</span> <span class="n">route_prob_max</span><span class="o">.</span><span class="n">detach</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>Multiply by the scaling factor</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">111</span> <span class="n">x</span> <span class="o">=</span> <span class="n">x</span> <span class="o">*</span> <span class="n">factor</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</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>Get indexes of tokens going to each expert</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">114</span> <span class="n">indexes_list</span> <span class="o">=</span> <span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">eq</span><span class="p">(</span><span class="n">routes</span><span class="p">,</span> <span class="n">i</span><span class="p">)</span><span class="o">.</span><span class="n">nonzero</span><span class="p">(</span><span class="n">as_tuple</span><span class="o">=</span><span class="kc">True</span><span class="p">)[</span><span class="mi">0</span><span class="p">]</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">n_experts</span><span class="p">)]</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>Initialize an empty tensor to store outputs</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">117</span> <span class="n">final_output</span> <span class="o">=</span> <span class="n">x</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">(</span><span class="n">x</span><span class="o">.</span><span class="n">shape</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>Capacity of each expert.
|
|
<script type="math/tex; mode=display">\mathrm{expert\;capacity} =
|
|
\frac{\mathrm{tokens\;per\;batch}}{\mathrm{number\;of\;experts}}
|
|
\times \mathrm{capacity\;factor}</script>
|
|
</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">123</span> <span class="n">capacity</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">capacity_factor</span> <span class="o">*</span> <span class="nb">len</span><span class="p">(</span><span class="n">x</span><span class="p">)</span> <span class="o">/</span> <span class="bp">self</span><span class="o">.</span><span class="n">n_experts</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-17'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-17'>#</a>
|
|
</div>
|
|
<p>Number of tokens routed to each expert.</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">125</span> <span class="n">counts</span> <span class="o">=</span> <span class="n">x</span><span class="o">.</span><span class="n">new_tensor</span><span class="p">([</span><span class="nb">len</span><span class="p">(</span><span class="n">indexes_list</span><span class="p">[</span><span class="n">i</span><span class="p">])</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">n_experts</span><span class="p">)])</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-18'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-18'>#</a>
|
|
</div>
|
|
<p>Initialize an empty list of dropped tokens</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">128</span> <span class="n">dropped</span> <span class="o">=</span> <span class="p">[]</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-19'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-19'>#</a>
|
|
</div>
|
|
<p>Only drop tokens if <code>drop_tokens</code> is <code>True</code>.</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">130</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">drop_tokens</span><span class="p">:</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-20'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-20'>#</a>
|
|
</div>
|
|
<p>Drop tokens in each of the experts</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">132</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">n_experts</span><span class="p">):</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-21'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-21'>#</a>
|
|
</div>
|
|
<p>Ignore if the expert is not over capacity</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">134</span> <span class="k">if</span> <span class="nb">len</span><span class="p">(</span><span class="n">indexes_list</span><span class="p">[</span><span class="n">i</span><span class="p">])</span> <span class="o"><=</span> <span class="n">capacity</span><span class="p">:</span>
|
|
<span class="lineno">135</span> <span class="k">continue</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-22'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-22'>#</a>
|
|
</div>
|
|
<p>Shuffle indexes before dropping</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">137</span> <span class="n">indexes_list</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">=</span> <span class="n">indexes_list</span><span class="p">[</span><span class="n">i</span><span class="p">][</span><span class="n">torch</span><span class="o">.</span><span class="n">randperm</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="n">indexes_list</span><span class="p">[</span><span class="n">i</span><span class="p">]))]</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-23'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-23'>#</a>
|
|
</div>
|
|
<p>Collect the tokens over capacity as dropped tokens</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">139</span> <span class="n">dropped</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">indexes_list</span><span class="p">[</span><span class="n">i</span><span class="p">][</span><span class="n">capacity</span><span class="p">:])</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-24'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-24'>#</a>
|
|
</div>
|
|
<p>Keep only the tokens upto the capacity of the expert</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">141</span> <span class="n">indexes_list</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">=</span> <span class="n">indexes_list</span><span class="p">[</span><span class="n">i</span><span class="p">][:</span><span class="n">capacity</span><span class="p">]</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-25'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-25'>#</a>
|
|
</div>
|
|
<p>Get outputs of the expert FFNs</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">144</span> <span class="n">route_outputs</span> <span class="o">=</span> <span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">experts</span><span class="p">[</span><span class="n">i</span><span class="p">](</span><span class="n">x</span><span class="p">[</span><span class="n">indexes_list</span><span class="p">[</span><span class="n">i</span><span class="p">],</span> <span class="p">:])</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">n_experts</span><span class="p">)]</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-26'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-26'>#</a>
|
|
</div>
|
|
<p>Assign to final output</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">147</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">n_experts</span><span class="p">):</span>
|
|
<span class="lineno">148</span> <span class="n">final_output</span><span class="p">[</span><span class="n">indexes_list</span><span class="p">[</span><span class="n">i</span><span class="p">],</span> <span class="p">:]</span> <span class="o">=</span> <span class="n">route_outputs</span><span class="p">[</span><span class="n">i</span><span class="p">]</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-27'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-27'>#</a>
|
|
</div>
|
|
<p>Pass through the dropped tokens</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">151</span> <span class="k">if</span> <span class="n">dropped</span><span class="p">:</span>
|
|
<span class="lineno">152</span> <span class="n">dropped</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">dropped</span><span class="p">)</span>
|
|
<span class="lineno">153</span> <span class="n">final_output</span><span class="p">[</span><span class="n">dropped</span><span class="p">,</span> <span class="p">:]</span> <span class="o">=</span> <span class="n">x</span><span class="p">[</span><span class="n">dropped</span><span class="p">,</span> <span class="p">:]</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-28'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-28'>#</a>
|
|
</div>
|
|
<p>Change the shape of the final output back to <code>[seq_len, batch_size, d_model]</code></p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">156</span> <span class="n">final_output</span> <span class="o">=</span> <span class="n">final_output</span><span class="o">.</span><span class="n">view</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> <span class="n">d_model</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-29'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-29'>#</a>
|
|
</div>
|
|
<p>Return
|
|
* the final output
|
|
* number of tokens routed to each expert
|
|
* sum of probabilities for each expert
|
|
* number of tokens dropped.
|
|
These are used for the load balancing loss and logging</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">164</span> <span class="k">return</span> <span class="n">final_output</span><span class="p">,</span> <span class="n">counts</span><span class="p">,</span> <span class="n">route_prob</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span> <span class="nb">len</span><span class="p">(</span><span class="n">dropped</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-30'>
|
|
<div class='docs doc-strings'>
|
|
<div class='section-link'>
|
|
<a href='#section-30'>#</a>
|
|
</div>
|
|
<h1>Switch Transformer Block</h1>
|
|
<p>This is the same as <a href="../models.html#TransformerLayer">normal transformer block</a>
|
|
with handling extra outputs of switch feedforward module.</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">167</span><span class="k">class</span> <span class="nc">SwitchTransformerLayer</span><span class="p">(</span><span class="n">Module</span><span class="p">):</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-31'>
|
|
<div class='docs doc-strings'>
|
|
<div class='section-link'>
|
|
<a href='#section-31'>#</a>
|
|
</div>
|
|
<ul>
|
|
<li><code>d_model</code> is the token embedding size</li>
|
|
<li><code>attn</code> is the attention module</li>
|
|
<li><code>feed_forward</code> is the feed forward module (which is the switching module in this case)</li>
|
|
<li><code>dropout_prob</code> is the probability of dropping out after self attention and FFN</li>
|
|
</ul>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">174</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">175</span> <span class="n">d_model</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
|
|
<span class="lineno">176</span> <span class="n">attn</span><span class="p">:</span> <span class="n">MultiHeadAttention</span><span class="p">,</span>
|
|
<span class="lineno">177</span> <span class="n">feed_forward</span><span class="p">:</span> <span class="n">SwitchFeedForward</span><span class="p">,</span>
|
|
<span class="lineno">178</span> <span class="n">dropout_prob</span><span class="p">:</span> <span class="nb">float</span><span class="p">):</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-32'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-32'>#</a>
|
|
</div>
|
|
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">185</span> <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
|
|
<span class="lineno">186</span> <span class="bp">self</span><span class="o">.</span><span class="n">size</span> <span class="o">=</span> <span class="n">d_model</span>
|
|
<span class="lineno">187</span> <span class="bp">self</span><span class="o">.</span><span class="n">attn</span> <span class="o">=</span> <span class="n">attn</span>
|
|
<span class="lineno">188</span> <span class="bp">self</span><span class="o">.</span><span class="n">feed_forward</span> <span class="o">=</span> <span class="n">feed_forward</span>
|
|
<span class="lineno">189</span> <span class="bp">self</span><span class="o">.</span><span class="n">dropout</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Dropout</span><span class="p">(</span><span class="n">dropout_prob</span><span class="p">)</span>
|
|
<span class="lineno">190</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm_self_attn</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">LayerNorm</span><span class="p">([</span><span class="n">d_model</span><span class="p">])</span>
|
|
<span class="lineno">191</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm_ff</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">LayerNorm</span><span class="p">([</span><span class="n">d_model</span><span class="p">])</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-33'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-33'>#</a>
|
|
</div>
|
|
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">193</span> <span class="k">def</span> <span class="nf">forward</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">194</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>
|
|
<span class="lineno">195</span> <span class="n">mask</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-34'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-34'>#</a>
|
|
</div>
|
|
<p>Normalize the vectors before doing self attention</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">197</span> <span class="n">z</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm_self_attn</span><span class="p">(</span><span class="n">x</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-35'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-35'>#</a>
|
|
</div>
|
|
<p>Run through self attention, i.e. keys and values are from self</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">199</span> <span class="n">self_attn</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">attn</span><span class="p">(</span><span class="n">query</span><span class="o">=</span><span class="n">z</span><span class="p">,</span> <span class="n">key</span><span class="o">=</span><span class="n">z</span><span class="p">,</span> <span class="n">value</span><span class="o">=</span><span class="n">z</span><span class="p">,</span> <span class="n">mask</span><span class="o">=</span><span class="n">mask</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-36'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-36'>#</a>
|
|
</div>
|
|
<p>Add the self attention results</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">201</span> <span class="n">x</span> <span class="o">=</span> <span class="n">x</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">dropout</span><span class="p">(</span><span class="n">self_attn</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-37'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-37'>#</a>
|
|
</div>
|
|
<p>Normalize for feed-forward</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">204</span> <span class="n">z</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm_ff</span><span class="p">(</span><span class="n">x</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-38'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-38'>#</a>
|
|
</div>
|
|
<p>Pass through the switching feed-forward network</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">206</span> <span class="n">ff</span><span class="p">,</span> <span class="n">counts</span><span class="p">,</span> <span class="n">route_prob</span><span class="p">,</span> <span class="n">n_dropped</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">feed_forward</span><span class="p">(</span><span class="n">z</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-39'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-39'>#</a>
|
|
</div>
|
|
<p>Add the feed-forward results back</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">208</span> <span class="n">x</span> <span class="o">=</span> <span class="n">x</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">dropout</span><span class="p">(</span><span class="n">ff</span><span class="p">)</span>
|
|
<span class="lineno">209</span>
|
|
<span class="lineno">210</span> <span class="k">return</span> <span class="n">x</span><span class="p">,</span> <span class="n">counts</span><span class="p">,</span> <span class="n">route_prob</span><span class="p">,</span> <span class="n">n_dropped</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-40'>
|
|
<div class='docs doc-strings'>
|
|
<div class='section-link'>
|
|
<a href='#section-40'>#</a>
|
|
</div>
|
|
<h2>Switch Transformer</h2>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">213</span><span class="k">class</span> <span class="nc">SwitchTransformer</span><span class="p">(</span><span class="n">Module</span><span class="p">):</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-41'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-41'>#</a>
|
|
</div>
|
|
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">218</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="n">layer</span><span class="p">:</span> <span class="n">SwitchTransformerLayer</span><span class="p">,</span> <span class="n">n_layers</span><span class="p">:</span> <span class="nb">int</span><span class="p">):</span>
|
|
<span class="lineno">219</span> <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-42'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-42'>#</a>
|
|
</div>
|
|
<p>Make copies of the transformer layer</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">221</span> <span class="bp">self</span><span class="o">.</span><span class="n">layers</span> <span class="o">=</span> <span class="n">clone_module_list</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">n_layers</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-43'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-43'>#</a>
|
|
</div>
|
|
<p>Final normalization layer</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">223</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">LayerNorm</span><span class="p">([</span><span class="n">layer</span><span class="o">.</span><span class="n">size</span><span class="p">])</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-44'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-44'>#</a>
|
|
</div>
|
|
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">225</span> <span class="k">def</span> <span class="nf">forward</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> <span class="n">mask</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-45'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-45'>#</a>
|
|
</div>
|
|
<p>Run through each transformer layer</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">227</span> <span class="n">counts</span><span class="p">,</span> <span class="n">route_prob</span><span class="p">,</span> <span class="n">n_dropped</span> <span class="o">=</span> <span class="p">[],</span> <span class="p">[],</span> <span class="p">[]</span>
|
|
<span class="lineno">228</span> <span class="k">for</span> <span class="n">layer</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">layers</span><span class="p">:</span>
|
|
<span class="lineno">229</span> <span class="n">x</span><span class="p">,</span> <span class="n">f</span><span class="p">,</span> <span class="n">p</span><span class="p">,</span> <span class="n">n_d</span> <span class="o">=</span> <span class="n">layer</span><span class="p">(</span><span class="n">x</span><span class="o">=</span><span class="n">x</span><span class="p">,</span> <span class="n">mask</span><span class="o">=</span><span class="n">mask</span><span class="p">)</span>
|
|
<span class="lineno">230</span> <span class="n">counts</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">f</span><span class="p">)</span>
|
|
<span class="lineno">231</span> <span class="n">route_prob</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">p</span><span class="p">)</span>
|
|
<span class="lineno">232</span> <span class="n">n_dropped</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">n_d</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-46'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-46'>#</a>
|
|
</div>
|
|
<p>Finally, normalize the vectors</p>
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">234</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm</span><span class="p">(</span><span class="n">x</span><span class="p">)</span></pre></div>
|
|
</div>
|
|
</div>
|
|
<div class='section' id='section-47'>
|
|
<div class='docs'>
|
|
<div class='section-link'>
|
|
<a href='#section-47'>#</a>
|
|
</div>
|
|
|
|
</div>
|
|
<div class='code'>
|
|
<div class="highlight"><pre><span class="lineno">236</span> <span class="k">return</span> <span class="n">x</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">stack</span><span class="p">(</span><span class="n">counts</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">stack</span><span class="p">(</span><span class="n">route_prob</span><span class="p">),</span> <span class="n">n_dropped</span></pre></div>
|
|
</div>
|
|
</div>
|
|
</div>
|
|
</div>
|
|
<script src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.4/MathJax.js?config=TeX-AMS_HTML">
|
|
</script>
|
|
<!-- MathJax configuration -->
|
|
<script type="text/x-mathjax-config">
|
|
MathJax.Hub.Config({
|
|
tex2jax: {
|
|
inlineMath: [ ['$','$'] ],
|
|
displayMath: [ ['$$','$$'] ],
|
|
processEscapes: true,
|
|
processEnvironments: true
|
|
},
|
|
// Center justify equations in code and markdown cells. Elsewhere
|
|
// we use CSS to left justify single line equations in code cells.
|
|
displayAlign: 'center',
|
|
"HTML-CSS": { fonts: ["TeX"] }
|
|
});
|
|
</script>
|
|
<script>
|
|
function handleImages() {
|
|
var images = document.querySelectorAll('p>img')
|
|
|
|
console.log(images);
|
|
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> |