mirror of
https://github.com/labmlai/annotated_deep_learning_paper_implementations.git
synced 2025-08-26 16:50:39 +08:00
AFT (#54)
This commit is contained in:
Varuna Jayasiri
@ -96,6 +96,7 @@ implementations.</p>
|
||||
<li><a href="transformers/switch/index.html">Switch Transformer</a></li>
|
||||
<li><a href="transformers/fast_weights/index.html">Fast Weights Transformer</a></li>
|
||||
<li><a href="transformers/fnet/index.html">FNet</a></li>
|
||||
<li><a href="transformers/aft/index.html">Attention Free Transformer</a></li>
|
||||
</ul>
|
||||
<h4>✨ <a href="recurrent_highway_networks/index.html">Recurrent Highway Networks</a></h4>
|
||||
<h4>✨ <a href="lstm/index.html">LSTM</a></h4>
|
||||
|
||||
@ -456,7 +456,7 @@
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/index.html</loc>
|
||||
<lastmod>2021-05-21T16:30:00+00:00</lastmod>
|
||||
<lastmod>2021-05-26T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
@ -638,7 +638,7 @@
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/transformers/models.html</loc>
|
||||
<lastmod>2021-02-02T16:30:00+00:00</lastmod>
|
||||
<lastmod>2021-06-02T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
@ -701,7 +701,7 @@
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/transformers/index.html</loc>
|
||||
<lastmod>2021-03-14T16:30:00+00:00</lastmod>
|
||||
<lastmod>2021-05-26T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
@ -720,13 +720,6 @@
|
||||
</url>
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/transformers/feedback/README.html</loc>
|
||||
<lastmod>2021-02-27T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/transformers/feedback/experiment.html</loc>
|
||||
<lastmod>2021-02-27T16:30:00+00:00</lastmod>
|
||||
@ -811,6 +804,20 @@
|
||||
</url>
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/transformers/aft/index.html</loc>
|
||||
<lastmod>2021-06-02T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/transformers/aft/experiment.html</loc>
|
||||
<lastmod>2021-06-02T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/transformers/mha.html</loc>
|
||||
<lastmod>2021-02-19T16:30:00+00:00</lastmod>
|
||||
@ -861,12 +868,19 @@
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/transformers/fnet/experiment.html</loc>
|
||||
<loc>https://nn.labml.ai/transformers/fnet/readme.html</loc>
|
||||
<lastmod>2021-05-26T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/transformers/fnet/experiment.html</loc>
|
||||
<lastmod>2021-06-02T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/transformers/xl/experiment.html</loc>
|
||||
<lastmod>2021-02-19T16:30:00+00:00</lastmod>
|
||||
|
||||
559
docs/transformers/aft/experiment.html
Normal file
559
docs/transformers/aft/experiment.html
Normal file
@ -0,0 +1,559 @@
|
||||
<!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 experiment trains an Attention Free Transformer (AFT) based model on Tiny Shakespeare dataset."/>
|
||||
|
||||
<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="Attention Free Transformer (AFT) Experiment"/>
|
||||
<meta name="twitter:description" content="This experiment trains an Attention Free Transformer (AFT) based model on Tiny Shakespeare dataset."/>
|
||||
<meta name="twitter:site" content="@labmlai"/>
|
||||
<meta name="twitter:creator" content="@labmlai"/>
|
||||
|
||||
<meta property="og:url" content="https://nn.labml.ai/transformers/aft/experiment.html"/>
|
||||
<meta property="og:title" content="Attention Free Transformer (AFT) Experiment"/>
|
||||
<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="Attention Free Transformer (AFT) Experiment"/>
|
||||
<meta property="og:description" content="This experiment trains an Attention Free Transformer (AFT) based model on Tiny Shakespeare dataset."/>
|
||||
|
||||
<title>Attention Free Transformer (AFT) Experiment</title>
|
||||
<link rel="shortcut icon" href="/icon.png"/>
|
||||
<link rel="stylesheet" href="../../pylit.css">
|
||||
<link rel="canonical" href="https://nn.labml.ai/transformers/aft/experiment.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">aft</a>
|
||||
</p>
|
||||
<p>
|
||||
|
||||
<a href="https://github.com/lab-ml/labml_nn/tree/master/labml_nn/transformers/aft/experiment.py">
|
||||
<img alt="Github"
|
||||
src="https://img.shields.io/github/stars/lab-ml/nn?style=social"
|
||||
style="max-width:100%;"/></a>
|
||||
<a href="https://join.slack.com/t/labforml/shared_invite/zt-egj9zvq9-Dl3hhZqobexgT7aVKnD14g/"
|
||||
rel="nofollow">
|
||||
<img alt="Join Slact"
|
||||
src="https://img.shields.io/badge/slack-chat-green.svg?logo=slack"
|
||||
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><a href="index.html">Attention Free Transformer (AFT)</a> Experiment</h1>
|
||||
<p>This is an annotated PyTorch experiment to train a <a href="index.html">AFT model</a>.</p>
|
||||
<p>This is based on
|
||||
<a href="../../experiments/nlp_autoregression.html">general training loop and configurations for auto-regressive NLP task</a>.</p>
|
||||
<p><a href="https://app.labml.ai/run/6348e504c3a511eba9529daa283fb495"><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">16</span><span></span><span class="kn">import</span> <span class="nn">torch</span>
|
||||
<span class="lineno">17</span>
|
||||
<span class="lineno">18</span><span class="kn">from</span> <span class="nn">labml</span> <span class="kn">import</span> <span class="n">experiment</span>
|
||||
<span class="lineno">19</span><span class="kn">from</span> <span class="nn">labml.configs</span> <span class="kn">import</span> <span class="n">option</span>
|
||||
<span class="lineno">20</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">21</span><span class="kn">from</span> <span class="nn">labml_nn.experiments.nlp_autoregression</span> <span class="kn">import</span> <span class="n">NLPAutoRegressionConfigs</span>
|
||||
<span class="lineno">22</span><span class="kn">from</span> <span class="nn">labml_nn.transformers</span> <span class="kn">import</span> <span class="n">TransformerConfigs</span><span class="p">,</span> <span class="n">Encoder</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>Simple autoregressive model</h2>
|
||||
<p>This consists of a token embedding layer, transformer encoder, and
|
||||
a final linear layer that gives token logits.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">25</span><span class="k">class</span> <span class="nc">AutoregressiveTransformer</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>encoder</code> is the transformer <a href="../models.html#Encoder">Encoder</a></li>
|
||||
<li><code>src_embed</code> is the token
|
||||
<a href="../models.html#EmbeddingsWithLearnedPositionalEncoding">embedding module (with positional encodings)</a></li>
|
||||
<li><code>generator</code> is the <a href="../models.html#Generator">final fully connected layer</a> that gives the logits.</li>
|
||||
</ul>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">33</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">encoder</span><span class="p">:</span> <span class="n">Encoder</span><span class="p">,</span> <span class="n">src_embed</span><span class="p">:</span> <span class="n">Module</span><span class="p">,</span> <span class="n">generator</span><span class="p">:</span> <span class="n">Module</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">40</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">41</span> <span class="bp">self</span><span class="o">.</span><span class="n">src_embed</span> <span class="o">=</span> <span class="n">src_embed</span>
|
||||
<span class="lineno">42</span> <span class="bp">self</span><span class="o">.</span><span class="n">encoder</span> <span class="o">=</span> <span class="n">encoder</span>
|
||||
<span class="lineno">43</span> <span class="bp">self</span><span class="o">.</span><span class="n">generator</span> <span class="o">=</span> <span class="n">generator</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-4'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-4'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">45</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-5'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-5'>#</a>
|
||||
</div>
|
||||
<p>Get the token embeddings with positional encodings</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">47</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">src_embed</span><span class="p">(</span><span class="n">x</span><span class="p">)</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>Transformer encoder</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">49</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">encoder</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="kc">None</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>Get logits</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">51</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">generator</span><span class="p">(</span><span class="n">x</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>Return results
|
||||
(second value is for state, since our trainer is used with RNNs also)</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">55</span> <span class="k">return</span> <span class="n">x</span><span class="p">,</span> <span class="kc">None</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-9'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-9'>#</a>
|
||||
</div>
|
||||
<h2>Configurations</h2>
|
||||
<p>This inherits from
|
||||
<a href="../../experiments/nlp_autoregression.html#NLPAutoRegressionConfigs"><code>NLPAutoRegressionConfigs</code></a></p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">58</span><span class="k">class</span> <span class="nc">Configs</span><span class="p">(</span><span class="n">NLPAutoRegressionConfigs</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>GPT model</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">67</span> <span class="n">model</span><span class="p">:</span> <span class="n">AutoregressiveTransformer</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>Transformer</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">69</span> <span class="n">transformer</span><span class="p">:</span> <span class="n">TransformerConfigs</span>
|
||||
<span class="lineno">70</span>
|
||||
<span class="lineno">71</span> <span class="n">local_window_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">32</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-12'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-12'>#</a>
|
||||
</div>
|
||||
<h3>Transformer configurations</h3>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">74</span><span class="nd">@option</span><span class="p">(</span><span class="n">Configs</span><span class="o">.</span><span class="n">transformer</span><span class="p">,</span> <span class="s1">'Transformer'</span><span class="p">)</span>
|
||||
<span class="lineno">75</span><span class="k">def</span> <span class="nf">_transformer_configs</span><span class="p">(</span><span class="n">c</span><span class="p">:</span> <span class="n">Configs</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>We use our
|
||||
<a href="../configs.html#TransformerConfigs">configurable transformer implementation</a></p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">82</span> <span class="n">conf</span> <span class="o">=</span> <span class="n">TransformerConfigs</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>Set the vocabulary sizes for embeddings and generating logits</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">84</span> <span class="n">conf</span><span class="o">.</span><span class="n">n_src_vocab</span> <span class="o">=</span> <span class="n">c</span><span class="o">.</span><span class="n">n_tokens</span>
|
||||
<span class="lineno">85</span> <span class="n">conf</span><span class="o">.</span><span class="n">n_tgt_vocab</span> <span class="o">=</span> <span class="n">c</span><span class="o">.</span><span class="n">n_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>Replace self-attention with an <a href="index.html">AFT Local Module</a></p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">87</span> <span class="kn">from</span> <span class="nn">labml_nn.transformers.aft</span> <span class="kn">import</span> <span class="n">AFTLocalAutoregressive</span>
|
||||
<span class="lineno">88</span> <span class="n">conf</span><span class="o">.</span><span class="n">encoder_attn</span> <span class="o">=</span> <span class="n">AFTLocalAutoregressive</span><span class="p">(</span><span class="n">c</span><span class="o">.</span><span class="n">d_model</span><span class="p">,</span> <span class="n">c</span><span class="o">.</span><span class="n">seq_len</span><span class="p">,</span> <span class="n">c</span><span class="o">.</span><span class="n">local_window_size</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>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">91</span> <span class="k">return</span> <span class="n">conf</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-17'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-17'>#</a>
|
||||
</div>
|
||||
<p>Create an auto-regressive model</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">94</span><span class="nd">@option</span><span class="p">(</span><span class="n">Configs</span><span class="o">.</span><span class="n">model</span><span class="p">)</span>
|
||||
<span class="lineno">95</span><span class="k">def</span> <span class="nf">_model</span><span class="p">(</span><span class="n">c</span><span class="p">:</span> <span class="n">Configs</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>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">99</span> <span class="n">m</span> <span class="o">=</span> <span class="n">AutoregressiveTransformer</span><span class="p">(</span><span class="n">c</span><span class="o">.</span><span class="n">transformer</span><span class="o">.</span><span class="n">encoder</span><span class="p">,</span>
|
||||
<span class="lineno">100</span> <span class="n">c</span><span class="o">.</span><span class="n">transformer</span><span class="o">.</span><span class="n">src_embed</span><span class="p">,</span>
|
||||
<span class="lineno">101</span> <span class="n">c</span><span class="o">.</span><span class="n">transformer</span><span class="o">.</span><span class="n">generator</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">c</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
|
||||
<span class="lineno">102</span>
|
||||
<span class="lineno">103</span> <span class="k">return</span> <span class="n">m</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-19'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-19'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">106</span><span class="k">def</span> <span class="nf">main</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>Create experiment</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">108</span> <span class="n">experiment</span><span class="o">.</span><span class="n">create</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s2">"aft"</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>Create configs</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">110</span> <span class="n">conf</span> <span class="o">=</span> <span class="n">Configs</span><span class="p">()</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>Override configurations</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">112</span> <span class="n">experiment</span><span class="o">.</span><span class="n">configs</span><span class="p">(</span><span class="n">conf</span><span class="p">,</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>Use character level tokenizer</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">114</span> <span class="s1">'tokenizer'</span><span class="p">:</span> <span class="s1">'character'</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>Prompt separator is blank</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">116</span> <span class="s1">'prompt_separator'</span><span class="p">:</span> <span class="s1">''</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>Starting prompt for sampling</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">118</span> <span class="s1">'prompt'</span><span class="p">:</span> <span class="s1">'It is '</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>Use Tiny Shakespeare dataset</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">120</span> <span class="s1">'text'</span><span class="p">:</span> <span class="s1">'tiny_shakespeare'</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>Use a context size of $128$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">123</span> <span class="s1">'seq_len'</span><span class="p">:</span> <span class="mi">256</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>Train for $32$ epochs</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">125</span> <span class="s1">'epochs'</span><span class="p">:</span> <span class="mi">128</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>Batch size $128$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">127</span> <span class="s1">'batch_size'</span><span class="p">:</span> <span class="mi">32</span><span class="p">,</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-30'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-30'>#</a>
|
||||
</div>
|
||||
<p>Switch between training and validation for $10$ times
|
||||
per epoch</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">130</span> <span class="s1">'inner_iterations'</span><span class="p">:</span> <span class="mi">10</span><span class="p">,</span>
|
||||
<span class="lineno">131</span>
|
||||
<span class="lineno">132</span> <span class="s1">'d_model'</span><span class="p">:</span> <span class="mi">128</span><span class="p">,</span>
|
||||
<span class="lineno">133</span> <span class="s1">'transformer.d_model'</span><span class="p">:</span> <span class="mi">128</span><span class="p">,</span>
|
||||
<span class="lineno">134</span> <span class="s1">'transformer.ffn.d_ff'</span><span class="p">:</span> <span class="mi">256</span><span class="p">,</span>
|
||||
<span class="lineno">135</span>
|
||||
<span class="lineno">136</span> <span class="s1">'optimizer.optimizer'</span><span class="p">:</span> <span class="s1">'Noam'</span><span class="p">,</span>
|
||||
<span class="lineno">137</span> <span class="s1">'optimizer.learning_rate'</span><span class="p">:</span> <span class="mf">1.</span><span class="p">,</span>
|
||||
<span class="lineno">138</span> <span class="p">})</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-31'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-31'>#</a>
|
||||
</div>
|
||||
<p>Set models for saving and loading</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">141</span> <span class="n">experiment</span><span class="o">.</span><span class="n">add_pytorch_models</span><span class="p">({</span><span class="s1">'model'</span><span class="p">:</span> <span class="n">conf</span><span class="o">.</span><span class="n">model</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>
|
||||
<p>Start the experiment</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">144</span> <span class="k">with</span> <span class="n">experiment</span><span class="o">.</span><span class="n">start</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>
|
||||
<p>Run training</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">146</span> <span class="n">conf</span><span class="o">.</span><span class="n">run</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>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">150</span><span class="k">if</span> <span class="vm">__name__</span> <span class="o">==</span> <span class="s1">'__main__'</span><span class="p">:</span>
|
||||
<span class="lineno">151</span> <span class="n">main</span><span class="p">()</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>
|
||||
555
docs/transformers/aft/index.html
Normal file
555
docs/transformers/aft/index.html
Normal file
@ -0,0 +1,555 @@
|
||||
<!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 the AFT (Attention Free 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="An Attention Free Transformer"/>
|
||||
<meta name="twitter:description" content="This is an annotated implementation/tutorial the AFT (Attention Free 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/aft/index.html"/>
|
||||
<meta property="og:title" content="An Attention Free 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="An Attention Free Transformer"/>
|
||||
<meta property="og:description" content="This is an annotated implementation/tutorial the AFT (Attention Free Transformer) in PyTorch."/>
|
||||
|
||||
<title>An Attention Free Transformer</title>
|
||||
<link rel="shortcut icon" href="/icon.png"/>
|
||||
<link rel="stylesheet" href="../../pylit.css">
|
||||
<link rel="canonical" href="https://nn.labml.ai/transformers/aft/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">aft</a>
|
||||
</p>
|
||||
<p>
|
||||
|
||||
<a href="https://github.com/lab-ml/labml_nn/tree/master/labml_nn/transformers/aft/__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://join.slack.com/t/labforml/shared_invite/zt-egj9zvq9-Dl3hhZqobexgT7aVKnD14g/"
|
||||
rel="nofollow">
|
||||
<img alt="Join Slact"
|
||||
src="https://img.shields.io/badge/slack-chat-green.svg?logo=slack"
|
||||
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>An Attention Free Transformer</h1>
|
||||
<p>This is a <a href="https://pytorch.org">PyTorch</a> implementation of the paper
|
||||
<a href="https://papers.labml.ai/paper/2105.14103">An Attention Free Transformer</a>.</p>
|
||||
<p>This paper replaces the <a href="../mha.html">self-attention layer</a> with a new efficient operation,
|
||||
that has memory complexity of $\mathcal{O}(Td)$, where $T$ is the sequence length
|
||||
and $d$ is the dimensionality of embeddings.</p>
|
||||
<p>The paper introduces AFT along with AFT-local and AFT-conv.
|
||||
Here we have implemented AFT-local which pays attention to closeby tokens
|
||||
in an autoregressive model.</p>
|
||||
<h2>Attention Free Transformer</h2>
|
||||
<p>AFT (similar to <a href="../mha.html">MHA</a>) first transforms the embeddings $X$ into
|
||||
query $Q = XW^Q$, key $K = XW^K$ and value $V = XW^V$ tensors with learned weights.
|
||||
The output for each position $t \in [1, T]$ is calculated with the following operation.</p>
|
||||
<p>
|
||||
<script type="math/tex; mode=display">Y_t = \sigma(Q_t) \odot
|
||||
\frac{\sum_{t'=1}^T \exp(K_{t'} + w_{t,t'}) \odot V_{t'}}
|
||||
{\sum_{t'=1}^T \exp(K_{t'} + w_{t,t'})}</script>
|
||||
</p>
|
||||
<p>, where $\odot$ is element-wise product, $\sigma$ is a non-linearity (sigmoid) and
|
||||
$w \in \mathbb{R}^{T \times T}$ is a learned matrix of pair-wise position biases.</p>
|
||||
<p>This means that we take the weighted average of values
|
||||
and multiply them by the query. This eliminates the need to calculate the $T \times T$ attention
|
||||
matrix that <a href="../mha.html">MHA</a> requires, and therefore reduce the memory requirement.</p>
|
||||
<h2>AFT Local</h2>
|
||||
<p>AFT Local only apply learned pair-wise position biases locally:</p>
|
||||
<p>
|
||||
<script type="math/tex; mode=display">\begin{align}
|
||||
w'_{t,t'} =
|
||||
\begin{cases}
|
||||
w_{t,t'}, & \text{for $\lvert t-t' \rvert \lt s$} \\
|
||||
0, & \text{otherwise}
|
||||
\end{cases}
|
||||
\end{align}</script>
|
||||
</p>
|
||||
<p>, where $s \le T$ is the local window size.</p>
|
||||
<p>Although $w’_{t,t’}$ is $0$ outside the local window the AFT operation still uses key-value pairs from
|
||||
other areas. This is different from local transformers where embeddings outside the local window are
|
||||
completely not visible.</p>
|
||||
<p>Here is <a href="experiment.html">the training code</a> for a AFT Local model.</p>
|
||||
<p><a href="https://app.labml.ai/run/6348e504c3a511eba9529daa283fb495"><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">61</span><span></span><span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">Optional</span>
|
||||
<span class="lineno">62</span>
|
||||
<span class="lineno">63</span><span class="kn">import</span> <span class="nn">torch</span>
|
||||
<span class="lineno">64</span><span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">nn</span>
|
||||
<span class="lineno">65</span>
|
||||
<span class="lineno">66</span><span class="kn">from</span> <span class="nn">labml_helpers.module</span> <span class="kn">import</span> <span class="n">Module</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>
|
||||
<h3>AFT Local Operation for Auto-Regression</h3>
|
||||
<p>This is an implementation of AFT Local for auto-regression, where $Y_t$
|
||||
only has visibility to tokens before $t$:</p>
|
||||
<p>
|
||||
<script type="math/tex; mode=display">Y_t = \sigma(Q_t) \odot
|
||||
\frac{\sum_{t'=1}^t \exp(K_{t'} + w_{t,t'}) \odot V_{t'}}
|
||||
{\sum_{t'=1}^t \exp(K_{t'} + w_{t,t'})}</script>
|
||||
</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">69</span><span class="k">class</span> <span class="nc">AFTLocalAutoregressive</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>d_model</code> is the number of features in the <code>query</code>, <code>key</code> and <code>value</code> vectors.</li>
|
||||
<li><code>seq_len</code> is $T$</li>
|
||||
<li><code>s</code> is the local window size $s$</li>
|
||||
<li><code>bias</code> is whether to have a bias parameter for transformations for $Q$, $K$ and $V$.</li>
|
||||
</ul>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">81</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">d_model</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span> <span class="n">seq_len</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span> <span class="n">s</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span> <span class="n">bias</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">True</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">89</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-4'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-4'>#</a>
|
||||
</div>
|
||||
<p>Local window size $s$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">92</span> <span class="bp">self</span><span class="o">.</span><span class="n">s</span> <span class="o">=</span> <span class="n">s</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>These transform the <code>query</code>, <code>key</code> and <code>value</code> vectors.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">94</span> <span class="bp">self</span><span class="o">.</span><span class="n">query</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">d_model</span><span class="p">,</span> <span class="n">bias</span><span class="o">=</span><span class="n">bias</span><span class="p">)</span>
|
||||
<span class="lineno">95</span> <span class="bp">self</span><span class="o">.</span><span class="n">key</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">d_model</span><span class="p">,</span> <span class="n">bias</span><span class="o">=</span><span class="n">bias</span><span class="p">)</span>
|
||||
<span class="lineno">96</span> <span class="bp">self</span><span class="o">.</span><span class="n">value</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">d_model</span><span class="p">,</span> <span class="n">bias</span><span class="o">=</span><span class="n">bias</span><span class="p">)</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>Pair-wise positional biases $w \in \mathbb{R}^{T \times T}$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">98</span> <span class="bp">self</span><span class="o">.</span><span class="n">pos_bias</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Parameter</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">seq_len</span><span class="p">,</span> <span class="n">seq_len</span><span class="p">),</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</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>Activation $\sigma$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">100</span> <span class="bp">self</span><span class="o">.</span><span class="n">activation</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Sigmoid</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>Output layer</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">102</span> <span class="bp">self</span><span class="o">.</span><span class="n">output</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">d_model</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-9'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-9'>#</a>
|
||||
</div>
|
||||
<p><code>query</code>, <code>key</code> and <code>value</code> are the tensors that store
|
||||
collection of token embeddings for <em>query</em>, <em>key</em> and <em>value</em>.
|
||||
They have shape <code>[seq_len, batch_size, d_model]</code>.</p>
|
||||
<p><code>mask</code> should be <code>None</code>. We keep this parameter so that we can use this as an
|
||||
drop in replacement for <a href="../mha.html">MHA</a>.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">104</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">105</span> <span class="n">query</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">106</span> <span class="n">key</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">107</span> <span class="n">value</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">108</span> <span class="n">mask</span><span class="p">:</span> <span class="n">Optional</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="o">=</span> <span class="kc">None</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><code>query</code>, <code>key</code> and <code>value</code> have shape <code>[seq_len, batch_size, d_model]</code></p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">119</span> <span class="n">seq_len</span><span class="p">,</span> <span class="n">_</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> <span class="n">query</span><span class="o">.</span><span class="n">shape</span>
|
||||
<span class="lineno">120</span>
|
||||
<span class="lineno">121</span> <span class="n">query</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">query</span><span class="p">(</span><span class="n">query</span><span class="p">)</span>
|
||||
<span class="lineno">122</span> <span class="n">key</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">key</span><span class="p">(</span><span class="n">key</span><span class="p">)</span>
|
||||
<span class="lineno">123</span> <span class="n">value</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">value</span><span class="p">(</span><span class="n">value</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>We subtract $\max(K_{t’} + w_{t,t’})$ before calculating the exponents to stabilize
|
||||
the softmax calculation.</p>
|
||||
<p>If $x_i$ is large $\exp(x_i)$ becomes huge and the computation of
|
||||
$\frac{\sum\exp(x_i)y_i}{\sum\exp(x_i)}$becomes unstable.
|
||||
Subtracting a constant before calculating the exponent from numerator and denominator will cancel out.
|
||||
and can help stabilize the computation.
|
||||
So we subtract $\max(x_i)$ to stabilize the computation.</p>
|
||||
<p>Here the maximum is the higher of $\max(K_{t’} + w_{t,t’})$ and $\max(K_{t’})$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">135</span> <span class="n">max_logit</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">max</span><span class="p">(</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>$\max(K_{t’})$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">137</span> <span class="n">key</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)[</span><span class="mi">0</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>$\max(K_{t’} + w_{t,t’})$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">139</span> <span class="p">(</span><span class="n">key</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">pos_bias</span><span class="p">[:</span><span class="n">seq_len</span><span class="p">,</span> <span class="p">:</span><span class="n">seq_len</span><span class="p">]</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)[</span><span class="mi">0</span><span class="p">]</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> <span class="mi">1</span><span class="p">))</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)[</span><span class="mi">0</span><span class="p">]</span>
|
||||
<span class="lineno">140</span> <span class="p">)[</span><span class="mi">0</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>
|
||||
<script type="math/tex; mode=display">\begin{align}
|
||||
Y_t &= \sigma(Q_t) \odot
|
||||
\frac{\sum_{t'=1}^t \exp(K_{t'} + w_{t,t'}) \odot V_{t'}}
|
||||
{\sum_{t'=1}^t \exp(K_{t'} + w_{t,t'})} \\
|
||||
&= \sigma(Q_t) \odot
|
||||
\frac{\sum_{t'=1}^{t-s} \exp(K_{t'}) \odot V_{t'} + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'}) \odot V_{t'}}
|
||||
{\sum_{t'=1}^{t-s} \exp(K_{t'}) + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'})} \\
|
||||
\end{align}</script>
|
||||
</p>
|
||||
<p>since
|
||||
<script type="math/tex; mode=display">\begin{align}
|
||||
w'_{t,t'} =
|
||||
\begin{cases}
|
||||
w_{t,t'}, & \text{for $\lvert t-t' \rvert \lt s$} \\
|
||||
0, & \text{otherwise}
|
||||
\end{cases}
|
||||
\end{align}</script>
|
||||
</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-15'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-15'>#</a>
|
||||
</div>
|
||||
<p>The numerator part $\sum_{t’=1}^{t-s} \exp(K_{t’}) \odot V_{t’}$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">162</span> <span class="n">num</span> <span class="o">=</span> <span class="n">key</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">(</span><span class="n">key</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</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>The denominator part $\sum_{t’=1}^{t-s} \exp(K_{t’})$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">164</span> <span class="n">den</span> <span class="o">=</span> <span class="n">key</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">(</span><span class="n">key</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</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>Output $Y$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">166</span> <span class="n">y</span> <span class="o">=</span> <span class="n">key</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">(</span><span class="n">key</span><span class="o">.</span><span class="n">shape</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>Iterate $t \in [0, T]$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">168</span> <span class="k">for</span> <span class="n">t</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">seq_len</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>$t - s + 1$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">170</span> <span class="n">f</span> <span class="o">=</span> <span class="n">t</span> <span class="o">-</span> <span class="bp">self</span><span class="o">.</span><span class="n">s</span> <span class="o">+</span> <span class="mi">1</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>This actually mean $t - s \ge 1$ since we are indexing from $1$ in the math equations and
|
||||
indexing from $0$ in code</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">173</span> <span class="k">if</span> <span class="n">f</span> <span class="o">>=</span> <span class="mi">1</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>$\exp(K_{t-s}$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">175</span> <span class="n">exp_l</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">exp</span><span class="p">(</span><span class="n">key</span><span class="p">[</span><span class="n">f</span> <span class="o">-</span> <span class="mi">1</span><span class="p">]</span> <span class="o">-</span> <span class="n">max_logit</span><span class="p">)</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>Update numerator and denominator parts</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">177</span> <span class="n">num</span> <span class="o">=</span> <span class="n">num</span> <span class="o">+</span> <span class="n">exp_l</span> <span class="o">*</span> <span class="n">value</span><span class="p">[</span><span class="n">f</span> <span class="o">-</span> <span class="mi">1</span><span class="p">]</span>
|
||||
<span class="lineno">178</span> <span class="n">den</span> <span class="o">=</span> <span class="n">den</span> <span class="o">+</span> <span class="n">exp_l</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>Start from the beginning if the local window size falls beyond</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">180</span> <span class="n">f</span> <span class="o">=</span> <span class="nb">max</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="n">f</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>$\exp(K_{t’} + w_{t,t’})$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">182</span> <span class="n">exp_l</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">exp</span><span class="p">(</span><span class="n">key</span><span class="p">[</span><span class="n">f</span><span class="p">:</span> <span class="n">t</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">pos_bias</span><span class="p">[</span><span class="n">t</span><span class="p">,</span> <span class="n">f</span><span class="p">:</span> <span class="n">t</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</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> <span class="mi">1</span><span class="p">)</span> <span class="o">-</span> <span class="n">max_logit</span><span class="o">.</span><span class="n">squeeze</span><span class="p">(</span><span class="mi">0</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>Numerator
|
||||
<script type="math/tex; mode=display">\sum_{t'=1}^{t-s} \exp(K_{t'}) \odot V_{t'} + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'}) \odot V_{t'}</script>
|
||||
</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">185</span> <span class="n">n</span> <span class="o">=</span> <span class="n">num</span> <span class="o">+</span> <span class="p">(</span><span class="n">exp_l</span> <span class="o">*</span> <span class="n">value</span><span class="p">[</span><span class="n">f</span><span class="p">:</span> <span class="n">t</span> <span class="o">+</span> <span class="mi">1</span><span class="p">])</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">0</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>Denominator
|
||||
<script type="math/tex; mode=display">\sum_{t'=1}^{t-s} \exp(K_{t'}) + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'})</script>
|
||||
</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">188</span> <span class="n">d</span> <span class="o">=</span> <span class="n">den</span> <span class="o">+</span> <span class="n">exp_l</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">0</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>
|
||||
<script type="math/tex; mode=display">Y_t = \sigma(Q_t) \odot
|
||||
\frac{\sum_{t'=1}^{t-s} \exp(K_{t'}) \odot V_{t'} + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'}) \odot V_{t'}}
|
||||
{\sum_{t'=1}^{t-s} \exp(K_{t'}) + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'})} </script>
|
||||
</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">192</span> <span class="n">y</span><span class="p">[</span><span class="n">t</span><span class="p">]</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">activation</span><span class="p">(</span><span class="n">query</span><span class="p">[</span><span class="n">t</span><span class="p">])</span> <span class="o">*</span> <span class="n">n</span> <span class="o">/</span> <span class="n">d</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>Output layer</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">195</span> <span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">output</span><span class="p">(</span><span class="n">y</span><span class="p">)</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>
|
||||
151
docs/transformers/aft/readme.html
Normal file
151
docs/transformers/aft/readme.html
Normal file
@ -0,0 +1,151 @@
|
||||
<!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=""/>
|
||||
|
||||
<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="An Attention Free Transformer"/>
|
||||
<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/aft/readme.html"/>
|
||||
<meta property="og:title" content="An Attention Free 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="An Attention Free Transformer"/>
|
||||
<meta property="og:description" content=""/>
|
||||
|
||||
<title>An Attention Free Transformer</title>
|
||||
<link rel="shortcut icon" href="/icon.png"/>
|
||||
<link rel="stylesheet" href="../../pylit.css">
|
||||
<link rel="canonical" href="https://nn.labml.ai/transformers/aft/readme.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">aft</a>
|
||||
</p>
|
||||
<p>
|
||||
|
||||
<a href="https://github.com/lab-ml/labml_nn/tree/master/labml_nn/transformers/aft/readme.md">
|
||||
<img alt="Github"
|
||||
src="https://img.shields.io/github/stars/lab-ml/nn?style=social"
|
||||
style="max-width:100%;"/></a>
|
||||
<a href="https://join.slack.com/t/labforml/shared_invite/zt-egj9zvq9-Dl3hhZqobexgT7aVKnD14g/"
|
||||
rel="nofollow">
|
||||
<img alt="Join Slact"
|
||||
src="https://img.shields.io/badge/slack-chat-green.svg?logo=slack"
|
||||
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'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-0'>#</a>
|
||||
</div>
|
||||
<h1><a href="https://nn.labml.ai/transformers/aft/index.html">An Attention Free Transformer</a></h1>
|
||||
<p>This is a <a href="https://pytorch.org">PyTorch</a> implementation of the paper
|
||||
<a href="https://papers.labml.ai/paper/2105.14103">An Attention Free Transformer</a>.</p>
|
||||
<p>This paper replaces the <a href="https://nn.labml.ai/transformers/mha.html">self-attention layer</a>
|
||||
with a new efficient operation,
|
||||
that has memory complexity of O(Td), where T is the sequence length
|
||||
and $d$ is the dimensionality of embeddings.</p>
|
||||
<p>The paper introduces AFT along with AFT-local and AFT-conv.
|
||||
Here we have implemented AFT-local which pays attention to closeby tokens
|
||||
in an autoregressive model.</p>
|
||||
<p><a href="https://app.labml.ai/run/6348e504c3a511eba9529daa283fb495"><img alt="View Run" src="https://img.shields.io/badge/labml-experiment-brightgreen" /></a></p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
|
||||
</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>
|
||||
@ -12,7 +12,7 @@
|
||||
<meta name="twitter:site" content="@labmlai"/>
|
||||
<meta name="twitter:creator" content="@labmlai"/>
|
||||
|
||||
<meta property="og:url" content="https://nn.labml.ai/transformers/feedback/README.html"/>
|
||||
<meta property="og:url" content="https://nn.labml.ai/transformers/feedback/readme.html"/>
|
||||
<meta property="og:title" content="Feedback 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"/>
|
||||
@ -23,7 +23,7 @@
|
||||
<title>Feedback Transformer</title>
|
||||
<link rel="shortcut icon" href="/icon.png"/>
|
||||
<link rel="stylesheet" href="../../pylit.css">
|
||||
<link rel="canonical" href="https://nn.labml.ai/transformers/feedback/README.html"/>
|
||||
<link rel="canonical" href="https://nn.labml.ai/transformers/feedback/readme.html"/>
|
||||
<!-- Global site tag (gtag.js) - Google Analytics -->
|
||||
<script async src="https://www.googletagmanager.com/gtag/js?id=G-4V3HC8HBLH"></script>
|
||||
<script>
|
||||
@ -50,7 +50,7 @@
|
||||
</p>
|
||||
<p>
|
||||
|
||||
<a href="https://github.com/lab-ml/labml_nn/tree/master/labml_nn/transformers/feedback/README.md">
|
||||
<a href="https://github.com/lab-ml/labml_nn/tree/master/labml_nn/transformers/feedback/readme.md">
|
||||
<img alt="Github"
|
||||
src="https://img.shields.io/github/stars/lab-ml/nn?style=social"
|
||||
style="max-width:100%;"/></a>
|
||||
|
||||
@ -421,10 +421,11 @@ per epoch</p>
|
||||
<div class='section-link'>
|
||||
<a href='#section-28'>#</a>
|
||||
</div>
|
||||
<p>Use <a href="index.html">FNet</a> instead of self-attention</p>
|
||||
<p>Use <a href="index.html">FNet</a> instead of self-a
|
||||
ttention</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">136</span> <span class="s1">'transformer.encoder_attn'</span><span class="p">:</span> <span class="s1">'fnet_mix'</span><span class="p">,</span></pre></div>
|
||||
<div class="highlight"><pre><span class="lineno">137</span> <span class="s1">'transformer.encoder_attn'</span><span class="p">:</span> <span class="s1">'fnet_mix'</span><span class="p">,</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-29'>
|
||||
@ -435,8 +436,9 @@ per epoch</p>
|
||||
<p>Use <a href="../../optimizers/noam.html">Noam optimizer</a></p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">139</span> <span class="s1">'optimizer.optimizer'</span><span class="p">:</span> <span class="s1">'Noam'</span><span class="p">,</span>
|
||||
<span class="lineno">140</span> <span class="p">})</span></pre></div>
|
||||
<div class="highlight"><pre><span class="lineno">140</span> <span class="s1">'optimizer.optimizer'</span><span class="p">:</span> <span class="s1">'Noam'</span><span class="p">,</span>
|
||||
<span class="lineno">141</span> <span class="s1">'optimizer.learning_rate'</span><span class="p">:</span> <span class="mf">1.</span><span class="p">,</span>
|
||||
<span class="lineno">142</span> <span class="p">})</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-30'>
|
||||
@ -447,7 +449,7 @@ per epoch</p>
|
||||
<p>Set models for saving and loading</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">143</span> <span class="n">experiment</span><span class="o">.</span><span class="n">add_pytorch_models</span><span class="p">({</span><span class="s1">'model'</span><span class="p">:</span> <span class="n">conf</span><span class="o">.</span><span class="n">model</span><span class="p">})</span></pre></div>
|
||||
<div class="highlight"><pre><span class="lineno">145</span> <span class="n">experiment</span><span class="o">.</span><span class="n">add_pytorch_models</span><span class="p">({</span><span class="s1">'model'</span><span class="p">:</span> <span class="n">conf</span><span class="o">.</span><span class="n">model</span><span class="p">})</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-31'>
|
||||
@ -458,7 +460,7 @@ per epoch</p>
|
||||
<p>Start the experiment</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">146</span> <span class="k">with</span> <span class="n">experiment</span><span class="o">.</span><span class="n">start</span><span class="p">():</span></pre></div>
|
||||
<div class="highlight"><pre><span class="lineno">148</span> <span class="k">with</span> <span class="n">experiment</span><span class="o">.</span><span class="n">start</span><span class="p">():</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-32'>
|
||||
@ -469,7 +471,7 @@ per epoch</p>
|
||||
<p>Run training</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">148</span> <span class="n">conf</span><span class="o">.</span><span class="n">run</span><span class="p">()</span></pre></div>
|
||||
<div class="highlight"><pre><span class="lineno">150</span> <span class="n">conf</span><span class="o">.</span><span class="n">run</span><span class="p">()</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-33'>
|
||||
@ -480,8 +482,8 @@ per epoch</p>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">152</span><span class="k">if</span> <span class="vm">__name__</span> <span class="o">==</span> <span class="s1">'__main__'</span><span class="p">:</span>
|
||||
<span class="lineno">153</span> <span class="n">main</span><span class="p">()</span></pre></div>
|
||||
<div class="highlight"><pre><span class="lineno">154</span><span class="k">if</span> <span class="vm">__name__</span> <span class="o">==</span> <span class="s1">'__main__'</span><span class="p">:</span>
|
||||
<span class="lineno">155</span> <span class="n">main</span><span class="p">()</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
@ -110,12 +110,15 @@ It does single GPU training but we implement the concept of switching as describ
|
||||
<h2><a href="fnet/index.html">FNet: Mixing Tokens with Fourier Transforms</a></h2>
|
||||
<p>This is an implementation of the paper
|
||||
<a href="https://arxiv.org/abs/2105.03824">FNet: Mixing Tokens with Fourier Transforms</a>.</p>
|
||||
<h2><a href="aft/index.html">Attention Free Transformer</a></h2>
|
||||
<p>This is an implementation of the paper
|
||||
<a href="https://papers.labml.ai/paper/2105.14103">An Attention Free Transformer</a>.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">67</span><span></span><span class="kn">from</span> <span class="nn">.configs</span> <span class="kn">import</span> <span class="n">TransformerConfigs</span>
|
||||
<span class="lineno">68</span><span class="kn">from</span> <span class="nn">.models</span> <span class="kn">import</span> <span class="n">TransformerLayer</span><span class="p">,</span> <span class="n">Encoder</span><span class="p">,</span> <span class="n">Decoder</span><span class="p">,</span> <span class="n">Generator</span><span class="p">,</span> <span class="n">EncoderDecoder</span>
|
||||
<span class="lineno">69</span><span class="kn">from</span> <span class="nn">.mha</span> <span class="kn">import</span> <span class="n">MultiHeadAttention</span>
|
||||
<span class="lineno">70</span><span class="kn">from</span> <span class="nn">labml_nn.transformers.xl.relative_mha</span> <span class="kn">import</span> <span class="n">RelativeMultiHeadAttention</span></pre></div>
|
||||
<div class="highlight"><pre><span class="lineno">72</span><span></span><span class="kn">from</span> <span class="nn">.configs</span> <span class="kn">import</span> <span class="n">TransformerConfigs</span>
|
||||
<span class="lineno">73</span><span class="kn">from</span> <span class="nn">.models</span> <span class="kn">import</span> <span class="n">TransformerLayer</span><span class="p">,</span> <span class="n">Encoder</span><span class="p">,</span> <span class="n">Decoder</span><span class="p">,</span> <span class="n">Generator</span><span class="p">,</span> <span class="n">EncoderDecoder</span>
|
||||
<span class="lineno">74</span><span class="kn">from</span> <span class="nn">.mha</span> <span class="kn">import</span> <span class="n">MultiHeadAttention</span>
|
||||
<span class="lineno">75</span><span class="kn">from</span> <span class="nn">labml_nn.transformers.xl.relative_mha</span> <span class="kn">import</span> <span class="n">RelativeMultiHeadAttention</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
@ -252,10 +252,10 @@ We found a detailed discussion about this in the paper
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">103</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">104</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">105</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>
|
||||
<span class="lineno">106</span> <span class="n">src</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
||||
<span class="lineno">107</span> <span class="n">src_mask</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">):</span></pre></div>
|
||||
<span class="lineno">104</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">105</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>
|
||||
<span class="lineno">106</span> <span class="n">src</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
||||
<span class="lineno">107</span> <span class="n">src_mask</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">):</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-12'>
|
||||
|
||||
@ -27,6 +27,7 @@ implementations.
|
||||
* [Switch Transformer](transformers/switch/index.html)
|
||||
* [Fast Weights Transformer](transformers/fast_weights/index.html)
|
||||
* [FNet](transformers/fnet/index.html)
|
||||
* [Attention Free Transformer](transformers/aft/index.html)
|
||||
|
||||
#### ✨ [Recurrent Highway Networks](recurrent_highway_networks/index.html)
|
||||
|
||||
|
||||
@ -62,6 +62,11 @@ This is an implementation of the paper
|
||||
|
||||
This is an implementation of the paper
|
||||
[FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824).
|
||||
|
||||
## [Attention Free Transformer](aft/index.html)
|
||||
|
||||
This is an implementation of the paper
|
||||
[An Attention Free Transformer](https://papers.labml.ai/paper/2105.14103).
|
||||
"""
|
||||
|
||||
from .configs import TransformerConfigs
|
||||
|
||||
195
labml_nn/transformers/aft/__init__.py
Normal file
195
labml_nn/transformers/aft/__init__.py
Normal file
@ -0,0 +1,195 @@
|
||||
"""
|
||||
---
|
||||
title: An Attention Free Transformer
|
||||
summary: >
|
||||
This is an annotated implementation/tutorial the AFT (Attention Free Transformer) in PyTorch.
|
||||
---
|
||||
|
||||
# An Attention Free Transformer
|
||||
|
||||
This is a [PyTorch](https://pytorch.org) implementation of the paper
|
||||
[An Attention Free Transformer](https://papers.labml.ai/paper/2105.14103).
|
||||
|
||||
This paper replaces the [self-attention layer](../mha.html) with a new efficient operation,
|
||||
that has memory complexity of $\mathcal{O}(Td)$, where $T$ is the sequence length
|
||||
and $d$ is the dimensionality of embeddings.
|
||||
|
||||
The paper introduces AFT along with AFT-local and AFT-conv.
|
||||
Here we have implemented AFT-local which pays attention to closeby tokens
|
||||
in an autoregressive model.
|
||||
|
||||
## Attention Free Transformer
|
||||
|
||||
AFT (similar to [MHA](../mha.html)) first transforms the embeddings $X$ into
|
||||
query $Q = XW^Q$, key $K = XW^K$ and value $V = XW^V$ tensors with learned weights.
|
||||
The output for each position $t \in [1, T]$ is calculated with the following operation.
|
||||
|
||||
$$Y_t = \sigma(Q_t) \odot
|
||||
\frac{\sum_{t'=1}^T \exp(K_{t'} + w_{t,t'}) \odot V_{t'}}
|
||||
{\sum_{t'=1}^T \exp(K_{t'} + w_{t,t'})}$$
|
||||
|
||||
, where $\odot$ is element-wise product, $\sigma$ is a non-linearity (sigmoid) and
|
||||
$w \in \mathbb{R}^{T \times T}$ is a learned matrix of pair-wise position biases.
|
||||
|
||||
This means that we take the weighted average of values
|
||||
and multiply them by the query. This eliminates the need to calculate the $T \times T$ attention
|
||||
matrix that [MHA](../mha.html) requires, and therefore reduce the memory requirement.
|
||||
|
||||
## AFT Local
|
||||
|
||||
AFT Local only apply learned pair-wise position biases locally:
|
||||
|
||||
\begin{align}
|
||||
w'_{t,t'} =
|
||||
\begin{cases}
|
||||
w_{t,t'}, & \text{for $\lvert t-t' \rvert \lt s$} \\
|
||||
0, & \text{otherwise}
|
||||
\end{cases}
|
||||
\end{align}
|
||||
|
||||
, where $s \le T$ is the local window size.
|
||||
|
||||
Although $w'_{t,t'}$ is $0$ outside the local window the AFT operation still uses key-value pairs from
|
||||
other areas. This is different from local transformers where embeddings outside the local window are
|
||||
completely not visible.
|
||||
|
||||
Here is [the training code](experiment.html) for a AFT Local model.
|
||||
|
||||
[](https://app.labml.ai/run/6348e504c3a511eba9529daa283fb495)
|
||||
"""
|
||||
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
|
||||
from labml_helpers.module import Module
|
||||
|
||||
|
||||
class AFTLocalAutoregressive(Module):
|
||||
"""
|
||||
### AFT Local Operation for Auto-Regression
|
||||
|
||||
This is an implementation of AFT Local for auto-regression, where $Y_t$
|
||||
only has visibility to tokens before $t$:
|
||||
|
||||
$$Y_t = \sigma(Q_t) \odot
|
||||
\frac{\sum_{t'=1}^t \exp(K_{t'} + w_{t,t'}) \odot V_{t'}}
|
||||
{\sum_{t'=1}^t \exp(K_{t'} + w_{t,t'})}$$
|
||||
"""
|
||||
|
||||
def __init__(self, d_model: int, seq_len: int, s: int, bias: bool = True):
|
||||
"""
|
||||
* `d_model` is the number of features in the `query`, `key` and `value` vectors.
|
||||
* `seq_len` is $T$
|
||||
* `s` is the local window size $s$
|
||||
* `bias` is whether to have a bias parameter for transformations for $Q$, $K$ and $V$.
|
||||
"""
|
||||
|
||||
super().__init__()
|
||||
|
||||
# Local window size $s$
|
||||
self.s = s
|
||||
# These transform the `query`, `key` and `value` vectors.
|
||||
self.query = nn.Linear(d_model, d_model, bias=bias)
|
||||
self.key = nn.Linear(d_model, d_model, bias=bias)
|
||||
self.value = nn.Linear(d_model, d_model, bias=bias)
|
||||
# Pair-wise positional biases $w \in \mathbb{R}^{T \times T}$
|
||||
self.pos_bias = nn.Parameter(torch.zeros(seq_len, seq_len), requires_grad=True)
|
||||
# Activation $\sigma$
|
||||
self.activation = nn.Sigmoid()
|
||||
# Output layer
|
||||
self.output = nn.Linear(d_model, d_model)
|
||||
|
||||
def forward(self, *,
|
||||
query: torch.Tensor,
|
||||
key: torch.Tensor,
|
||||
value: torch.Tensor,
|
||||
mask: Optional[torch.Tensor] = None):
|
||||
"""
|
||||
`query`, `key` and `value` are the tensors that store
|
||||
collection of token embeddings for *query*, *key* and *value*.
|
||||
They have shape `[seq_len, batch_size, d_model]`.
|
||||
|
||||
`mask` should be `None`. We keep this parameter so that we can use this as an
|
||||
drop in replacement for [MHA](../mha.html).
|
||||
"""
|
||||
|
||||
# `query`, `key` and `value` have shape `[seq_len, batch_size, d_model]`
|
||||
seq_len, _, _ = query.shape
|
||||
|
||||
query = self.query(query)
|
||||
key = self.key(key)
|
||||
value = self.value(value)
|
||||
|
||||
# We subtract $\max(K_{t'} + w_{t,t'})$ before calculating the exponents to stabilize
|
||||
# the softmax calculation.
|
||||
#
|
||||
# If $x_i$ is large $\exp(x_i)$ becomes huge and the computation of
|
||||
# $\frac{\sum\exp(x_i)y_i}{\sum\exp(x_i)}$becomes unstable.
|
||||
# Subtracting a constant before calculating the exponent from numerator and denominator will cancel out.
|
||||
# and can help stabilize the computation.
|
||||
# So we subtract $\max(x_i)$ to stabilize the computation.
|
||||
#
|
||||
# Here the maximum is the higher of $\max(K_{t'} + w_{t,t'})$ and $\max(K_{t'})$
|
||||
max_logit = torch.max(
|
||||
# $\max(K_{t'})$
|
||||
key.max(dim=0)[0],
|
||||
# $\max(K_{t'} + w_{t,t'})$
|
||||
(key + self.pos_bias[:seq_len, :seq_len].max(dim=0)[0].view(-1, 1, 1)).max(dim=0)[0]
|
||||
)[0]
|
||||
|
||||
# \begin{align}
|
||||
# Y_t &= \sigma(Q_t) \odot
|
||||
# \frac{\sum_{t'=1}^t \exp(K_{t'} + w_{t,t'}) \odot V_{t'}}
|
||||
# {\sum_{t'=1}^t \exp(K_{t'} + w_{t,t'})} \\
|
||||
# &= \sigma(Q_t) \odot
|
||||
# \frac{\sum_{t'=1}^{t-s} \exp(K_{t'}) \odot V_{t'} + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'}) \odot V_{t'}}
|
||||
# {\sum_{t'=1}^{t-s} \exp(K_{t'}) + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'})} \\
|
||||
# \end{align}
|
||||
#
|
||||
# since
|
||||
# \begin{align}
|
||||
# w'_{t,t'} =
|
||||
# \begin{cases}
|
||||
# w_{t,t'}, & \text{for $\lvert t-t' \rvert \lt s$} \\
|
||||
# 0, & \text{otherwise}
|
||||
# \end{cases}
|
||||
# \end{align}
|
||||
#
|
||||
|
||||
# The numerator part $\sum_{t'=1}^{t-s} \exp(K_{t'}) \odot V_{t'}$
|
||||
num = key.new_zeros(key.shape[1:])
|
||||
# The denominator part $\sum_{t'=1}^{t-s} \exp(K_{t'})$
|
||||
den = key.new_zeros(key.shape[1:])
|
||||
# Output $Y$
|
||||
y = key.new_zeros(key.shape)
|
||||
# Iterate $t \in [0, T]$
|
||||
for t in range(seq_len):
|
||||
# $t - s + 1$
|
||||
f = t - self.s + 1
|
||||
# This actually mean $t - s \ge 1$ since we are indexing from $1$ in the math equations and
|
||||
# indexing from $0$ in code
|
||||
if f >= 1:
|
||||
# $\exp(K_{t-s}$
|
||||
exp_l = torch.exp(key[f - 1] - max_logit)
|
||||
# Update numerator and denominator parts
|
||||
num = num + exp_l * value[f - 1]
|
||||
den = den + exp_l
|
||||
# Start from the beginning if the local window size falls beyond
|
||||
f = max(0, f)
|
||||
# $\exp(K_{t'} + w_{t,t'})$
|
||||
exp_l = torch.exp(key[f: t + 1] + self.pos_bias[t, f: t + 1].view(-1, 1, 1) - max_logit.squeeze(0))
|
||||
# Numerator
|
||||
# $$\sum_{t'=1}^{t-s} \exp(K_{t'}) \odot V_{t'} + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'}) \odot V_{t'}$$
|
||||
n = num + (exp_l * value[f: t + 1]).sum(dim=0)
|
||||
# Denominator
|
||||
# $$\sum_{t'=1}^{t-s} \exp(K_{t'}) + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'})$$
|
||||
d = den + exp_l.sum(dim=0)
|
||||
# $$Y_t = \sigma(Q_t) \odot
|
||||
# \frac{\sum_{t'=1}^{t-s} \exp(K_{t'}) \odot V_{t'} + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'}) \odot V_{t'}}
|
||||
# {\sum_{t'=1}^{t-s} \exp(K_{t'}) + \sum_{t'=t-s+1}^t \exp(K_{t'} + w_{t,t'})} $$
|
||||
y[t] = self.activation(query[t]) * n / d
|
||||
|
||||
# Output layer
|
||||
return self.output(y)
|
||||
151
labml_nn/transformers/aft/experiment.py
Normal file
151
labml_nn/transformers/aft/experiment.py
Normal file
@ -0,0 +1,151 @@
|
||||
"""
|
||||
---
|
||||
title: Attention Free Transformer (AFT) Experiment
|
||||
summary: This experiment trains an Attention Free Transformer (AFT) based model on Tiny Shakespeare dataset.
|
||||
---
|
||||
|
||||
# [Attention Free Transformer (AFT)](index.html) Experiment
|
||||
|
||||
This is an annotated PyTorch experiment to train a [AFT model](index.html).
|
||||
|
||||
This is based on
|
||||
[general training loop and configurations for auto-regressive NLP task](../../experiments/nlp_autoregression.html).
|
||||
|
||||
[](https://app.labml.ai/run/6348e504c3a511eba9529daa283fb495)
|
||||
"""
|
||||
import torch
|
||||
|
||||
from labml import experiment
|
||||
from labml.configs import option
|
||||
from labml_helpers.module import Module
|
||||
from labml_nn.experiments.nlp_autoregression import NLPAutoRegressionConfigs
|
||||
from labml_nn.transformers import TransformerConfigs, Encoder
|
||||
|
||||
|
||||
class AutoregressiveTransformer(Module):
|
||||
"""
|
||||
## Simple autoregressive model
|
||||
|
||||
This consists of a token embedding layer, transformer encoder, and
|
||||
a final linear layer that gives token logits.
|
||||
"""
|
||||
|
||||
def __init__(self, encoder: Encoder, src_embed: Module, generator: Module):
|
||||
"""
|
||||
* `encoder` is the transformer [Encoder](../models.html#Encoder)
|
||||
* `src_embed` is the token
|
||||
[embedding module (with positional encodings)](../models.html#EmbeddingsWithLearnedPositionalEncoding)
|
||||
* `generator` is the [final fully connected layer](../models.html#Generator) that gives the logits.
|
||||
"""
|
||||
super().__init__()
|
||||
self.src_embed = src_embed
|
||||
self.encoder = encoder
|
||||
self.generator = generator
|
||||
|
||||
def forward(self, x: torch.Tensor):
|
||||
# Get the token embeddings with positional encodings
|
||||
x = self.src_embed(x)
|
||||
# Transformer encoder
|
||||
x = self.encoder(x, None)
|
||||
# Get logits
|
||||
x = self.generator(x)
|
||||
|
||||
# Return results
|
||||
# (second value is for state, since our trainer is used with RNNs also)
|
||||
return x, None
|
||||
|
||||
|
||||
class Configs(NLPAutoRegressionConfigs):
|
||||
"""
|
||||
## Configurations
|
||||
|
||||
This inherits from
|
||||
[`NLPAutoRegressionConfigs`](../../experiments/nlp_autoregression.html#NLPAutoRegressionConfigs)
|
||||
"""
|
||||
|
||||
# GPT model
|
||||
model: AutoregressiveTransformer
|
||||
# Transformer
|
||||
transformer: TransformerConfigs
|
||||
|
||||
local_window_size: int = 32
|
||||
|
||||
|
||||
@option(Configs.transformer, 'Transformer')
|
||||
def _transformer_configs(c: Configs):
|
||||
"""
|
||||
### Transformer configurations
|
||||
"""
|
||||
|
||||
# We use our
|
||||
# [configurable transformer implementation](../configs.html#TransformerConfigs)
|
||||
conf = TransformerConfigs()
|
||||
# Set the vocabulary sizes for embeddings and generating logits
|
||||
conf.n_src_vocab = c.n_tokens
|
||||
conf.n_tgt_vocab = c.n_tokens
|
||||
# Replace self-attention with an [AFT Local Module](index.html)
|
||||
from labml_nn.transformers.aft import AFTLocalAutoregressive
|
||||
conf.encoder_attn = AFTLocalAutoregressive(c.d_model, c.seq_len, c.local_window_size)
|
||||
|
||||
#
|
||||
return conf
|
||||
|
||||
|
||||
@option(Configs.model)
|
||||
def _model(c: Configs):
|
||||
"""
|
||||
Create an auto-regressive model
|
||||
"""
|
||||
m = AutoregressiveTransformer(c.transformer.encoder,
|
||||
c.transformer.src_embed,
|
||||
c.transformer.generator).to(c.device)
|
||||
|
||||
return m
|
||||
|
||||
|
||||
def main():
|
||||
# Create experiment
|
||||
experiment.create(name="aft")
|
||||
# Create configs
|
||||
conf = Configs()
|
||||
# Override configurations
|
||||
experiment.configs(conf, {
|
||||
# Use character level tokenizer
|
||||
'tokenizer': 'character',
|
||||
# Prompt separator is blank
|
||||
'prompt_separator': '',
|
||||
# Starting prompt for sampling
|
||||
'prompt': 'It is ',
|
||||
# Use Tiny Shakespeare dataset
|
||||
'text': 'tiny_shakespeare',
|
||||
|
||||
# Use a context size of $128$
|
||||
'seq_len': 256,
|
||||
# Train for $32$ epochs
|
||||
'epochs': 128,
|
||||
# Batch size $128$
|
||||
'batch_size': 32,
|
||||
# Switch between training and validation for $10$ times
|
||||
# per epoch
|
||||
'inner_iterations': 10,
|
||||
|
||||
'd_model': 128,
|
||||
'transformer.d_model': 128,
|
||||
'transformer.ffn.d_ff': 256,
|
||||
|
||||
'optimizer.optimizer': 'Noam',
|
||||
'optimizer.learning_rate': 1.,
|
||||
})
|
||||
|
||||
# Set models for saving and loading
|
||||
experiment.add_pytorch_models({'model': conf.model})
|
||||
|
||||
# Start the experiment
|
||||
with experiment.start():
|
||||
# Run training
|
||||
conf.run()
|
||||
|
||||
|
||||
#
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
15
labml_nn/transformers/aft/readme.md
Normal file
15
labml_nn/transformers/aft/readme.md
Normal file
@ -0,0 +1,15 @@
|
||||
# [An Attention Free Transformer](https://nn.labml.ai/transformers/aft/index.html)
|
||||
|
||||
This is a [PyTorch](https://pytorch.org) implementation of the paper
|
||||
[An Attention Free Transformer](https://papers.labml.ai/paper/2105.14103).
|
||||
|
||||
This paper replaces the [self-attention layer](https://nn.labml.ai/transformers/mha.html)
|
||||
with a new efficient operation,
|
||||
that has memory complexity of O(Td), where T is the sequence length
|
||||
and $d$ is the dimensionality of embeddings.
|
||||
|
||||
The paper introduces AFT along with AFT-local and AFT-conv.
|
||||
Here we have implemented AFT-local which pays attention to closeby tokens
|
||||
in an autoregressive model.
|
||||
|
||||
[](https://app.labml.ai/run/6348e504c3a511eba9529daa283fb495)
|
||||
@ -132,11 +132,13 @@ def main():
|
||||
'transformer.n_heads': 8,
|
||||
'transformer.n_layers': 6,
|
||||
|
||||
# Use [FNet](index.html) instead of self-attention
|
||||
# Use [FNet](index.html) instead of self-a
|
||||
# ttention
|
||||
'transformer.encoder_attn': 'fnet_mix',
|
||||
|
||||
# Use [Noam optimizer](../../optimizers/noam.html)
|
||||
'optimizer.optimizer': 'Noam',
|
||||
'optimizer.learning_rate': 1.,
|
||||
})
|
||||
|
||||
# Set models for saving and loading
|
||||
|
||||
@ -101,10 +101,10 @@ class TransformerLayer(Module):
|
||||
self.is_save_ff_input = False
|
||||
|
||||
def forward(self, *,
|
||||
x: torch.Tensor,
|
||||
mask: torch.Tensor,
|
||||
src: torch.Tensor = None,
|
||||
src_mask: torch.Tensor = None):
|
||||
x: torch.Tensor,
|
||||
mask: torch.Tensor,
|
||||
src: torch.Tensor = None,
|
||||
src_mask: torch.Tensor = None):
|
||||
# Normalize the vectors before doing self attention
|
||||
z = self.norm_self_attn(x)
|
||||
# Run through self attention, i.e. keys and values are from self
|
||||
|
||||
@ -33,6 +33,7 @@ implementations almost weekly.
|
||||
* [Switch Transformer](https://nn.labml.ai/transformers/switch/index.html)
|
||||
* [Fast Weights Transformer](https://nn.labml.ai/transformers/fast_weights/index.html)
|
||||
* [FNet](https://nn.labml.ai/transformers/fnet/index.html)
|
||||
* [Attention Free Transformer](https://nn.labml.ai/transformers/aft/index.html)
|
||||
|
||||
#### ✨ [Recurrent Highway Networks](https://nn.labml.ai/recurrent_highway_networks/index.html)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user