mirror of
https://github.com/labmlai/annotated_deep_learning_paper_implementations.git
synced 2025-08-16 10:51:23 +08:00
Conv mixer (#100)
This commit is contained in:
Varuna Jayasiri
BIN
docs/conv_mixer/conv_mixer.png
Normal file
BIN
docs/conv_mixer/conv_mixer.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 170 KiB |
377
docs/conv_mixer/experiment.html
Normal file
377
docs/conv_mixer/experiment.html
Normal file
@ -0,0 +1,377 @@
|
||||
<!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="Train ConvMixer on CIFAR 10"/>
|
||||
|
||||
<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="Train ConvMixer on CIFAR 10"/>
|
||||
<meta name="twitter:description" content="Train ConvMixer on CIFAR 10"/>
|
||||
<meta name="twitter:site" content="@labmlai"/>
|
||||
<meta name="twitter:creator" content="@labmlai"/>
|
||||
|
||||
<meta property="og:url" content="https://nn.labml.ai/conv_mixer/experiment.html"/>
|
||||
<meta property="og:title" content="Train ConvMixer on CIFAR 10"/>
|
||||
<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="Train ConvMixer on CIFAR 10"/>
|
||||
<meta property="og:description" content="Train ConvMixer on CIFAR 10"/>
|
||||
|
||||
<title>Train ConvMixer on CIFAR 10</title>
|
||||
<link rel="shortcut icon" href="/icon.png"/>
|
||||
<link rel="stylesheet" href="../pylit.css">
|
||||
<link rel="canonical" href="https://nn.labml.ai/conv_mixer/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">conv_mixer</a>
|
||||
</p>
|
||||
<p>
|
||||
|
||||
<a href="https://github.com/labmlai/annotated_deep_learning_paper_implementations/tree/master/labml_nn/conv_mixer/experiment.py">
|
||||
<img alt="Github"
|
||||
src="https://img.shields.io/github/stars/labmlai/annotated_deep_learning_paper_implementations?style=social"
|
||||
style="max-width:100%;"/></a>
|
||||
<a href="https://twitter.com/labmlai"
|
||||
rel="nofollow">
|
||||
<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>Train a <a href="index.html">ConvMixer</a> on CIFAR 10</h1>
|
||||
<p>This script trains a ConvMixer on CIFAR 10 dataset.</p>
|
||||
<p>This is not an attempt to reproduce the results of the paper.
|
||||
The paper uses image augmentations
|
||||
present in <a href="https://github.com/rwightman/pytorch-image-models">PyTorch Image Models (timm)</a>
|
||||
for training. We haven’t done this for simplicity - which causes our validation accuracy to drop.</p>
|
||||
<p><a href="https://app.labml.ai/run/0fc344da2cd011ecb0bc3fdb2e774a3d"><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">20</span><span></span><span class="kn">from</span> <span class="nn">labml</span> <span class="kn">import</span> <span class="n">experiment</span>
|
||||
<span class="lineno">21</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">22</span><span class="kn">from</span> <span class="nn">labml_nn.experiments.cifar10</span> <span class="kn">import</span> <span class="n">CIFAR10Configs</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>Configurations</h2>
|
||||
<p>We use <a href="../experiments/cifar10.html"><code>CIFAR10Configs</code></a> which defines all the
|
||||
dataset related configurations, optimizer, and a training loop.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">25</span><span class="k">class</span> <span class="nc">Configs</span><span class="p">(</span><span class="n">CIFAR10Configs</span><span class="p">):</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-2'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-2'>#</a>
|
||||
</div>
|
||||
<p>Size of a patch, $p$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">34</span> <span class="n">patch_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">2</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-3'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-3'>#</a>
|
||||
</div>
|
||||
<p>Number of channels in patch embeddings, $h$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">36</span> <span class="n">d_model</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">256</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>Number of <a href="#ConvMixerLayer">ConvMixer layers</a> or depth, $d$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">38</span> <span class="n">n_layers</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">8</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>Kernel size of the depth-wise convolution, $k$</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">40</span> <span class="n">kernel_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">7</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>Number of classes in the task</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">42</span> <span class="n">n_classes</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">10</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-7'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-7'>#</a>
|
||||
</div>
|
||||
<h3>Create model</h3>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">45</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">46</span><span class="k">def</span> <span class="nf">_conv_mixer</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-8'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-8'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">50</span> <span class="kn">from</span> <span class="nn">labml_nn.conv_mixer</span> <span class="kn">import</span> <span class="n">ConvMixerLayer</span><span class="p">,</span> <span class="n">ConvMixer</span><span class="p">,</span> <span class="n">ClassificationHead</span><span class="p">,</span> <span class="n">PatchEmbeddings</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-9'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-9'>#</a>
|
||||
</div>
|
||||
<p>Create ConvMixer</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">53</span> <span class="k">return</span> <span class="n">ConvMixer</span><span class="p">(</span><span class="n">ConvMixerLayer</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">kernel_size</span><span class="p">),</span> <span class="n">c</span><span class="o">.</span><span class="n">n_layers</span><span class="p">,</span>
|
||||
<span class="lineno">54</span> <span class="n">PatchEmbeddings</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">patch_size</span><span class="p">,</span> <span class="mi">3</span><span class="p">),</span>
|
||||
<span class="lineno">55</span> <span class="n">ClassificationHead</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">n_classes</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></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-10'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-10'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">58</span><span class="k">def</span> <span class="nf">main</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>Create experiment</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">60</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="s1">'ConvMixer'</span><span class="p">,</span> <span class="n">comment</span><span class="o">=</span><span class="s1">'cifar10'</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>Create configurations</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">62</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-13'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-13'>#</a>
|
||||
</div>
|
||||
<p>Load configurations</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">64</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-14'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-14'>#</a>
|
||||
</div>
|
||||
<p>Optimizer</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">66</span> <span class="s1">'optimizer.optimizer'</span><span class="p">:</span> <span class="s1">'Adam'</span><span class="p">,</span>
|
||||
<span class="lineno">67</span> <span class="s1">'optimizer.learning_rate'</span><span class="p">:</span> <span class="mf">2.5e-4</span><span class="p">,</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-15'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-15'>#</a>
|
||||
</div>
|
||||
<p>Training epochs and batch size</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">70</span> <span class="s1">'epochs'</span><span class="p">:</span> <span class="mi">150</span><span class="p">,</span>
|
||||
<span class="lineno">71</span> <span class="s1">'train_batch_size'</span><span class="p">:</span> <span class="mi">64</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>Simple image augmentations</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">74</span> <span class="s1">'train_dataset'</span><span class="p">:</span> <span class="s1">'cifar10_train_augmented'</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>Do not augment images for validation</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">76</span> <span class="s1">'valid_dataset'</span><span class="p">:</span> <span class="s1">'cifar10_valid_no_augment'</span><span class="p">,</span>
|
||||
<span class="lineno">77</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>Set model for saving/loading</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">79</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-19'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-19'>#</a>
|
||||
</div>
|
||||
<p>Start the experiment and run the training loop</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">81</span> <span class="k">with</span> <span class="n">experiment</span><span class="o">.</span><span class="n">start</span><span class="p">():</span>
|
||||
<span class="lineno">82</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-20'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-20'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">86</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">87</span> <span class="n">main</span><span class="p">()</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='footer'>
|
||||
<a href="https://papers.labml.ai">Trending Research Papers</a>
|
||||
<a href="https://labml.ai">labml.ai</a>
|
||||
</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>
|
||||
715
docs/conv_mixer/index.html
Normal file
715
docs/conv_mixer/index.html
Normal file
@ -0,0 +1,715 @@
|
||||
<!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="A PyTorch implementation/tutorial of the paper "Patches Are All You Need?""/>
|
||||
|
||||
<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="Patches Are All You Need? (ConvMixer)"/>
|
||||
<meta name="twitter:description" content="A PyTorch implementation/tutorial of the paper "Patches Are All You Need?""/>
|
||||
<meta name="twitter:site" content="@labmlai"/>
|
||||
<meta name="twitter:creator" content="@labmlai"/>
|
||||
|
||||
<meta property="og:url" content="https://nn.labml.ai/conv_mixer/index.html"/>
|
||||
<meta property="og:title" content="Patches Are All You Need? (ConvMixer)"/>
|
||||
<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="Patches Are All You Need? (ConvMixer)"/>
|
||||
<meta property="og:description" content="A PyTorch implementation/tutorial of the paper "Patches Are All You Need?""/>
|
||||
|
||||
<title>Patches Are All You Need? (ConvMixer)</title>
|
||||
<link rel="shortcut icon" href="/icon.png"/>
|
||||
<link rel="stylesheet" href="../pylit.css">
|
||||
<link rel="canonical" href="https://nn.labml.ai/conv_mixer/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">conv_mixer</a>
|
||||
</p>
|
||||
<p>
|
||||
|
||||
<a href="https://github.com/labmlai/annotated_deep_learning_paper_implementations/tree/master/labml_nn/conv_mixer/__init__.py">
|
||||
<img alt="Github"
|
||||
src="https://img.shields.io/github/stars/labmlai/annotated_deep_learning_paper_implementations?style=social"
|
||||
style="max-width:100%;"/></a>
|
||||
<a href="https://twitter.com/labmlai"
|
||||
rel="nofollow">
|
||||
<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>Patches Are All You Need? (ConvMixer)</h1>
|
||||
<p>This is a <a href="https://pytorch.org">PyTorch</a> implementation of the paper
|
||||
<a href="https://papers.labml.ai/paper/TVHS5Y4dNvM">Patches Are All You Need?</a>.</p>
|
||||
<p><img alt="ConvMixer diagram from the paper" src="conv_mixer.png" /></p>
|
||||
<p>ConvMixer is Similar to <a href="../transformers/mlp_mixer/index.html">MLP-Mixer</a>.
|
||||
MLP-Mixer separates mixing of spatial and channel dimensions, by applying a MLP across spatial dimension
|
||||
and then an MLP across the channel dimension
|
||||
(spatial MLP replaces the <a href="../transformers/vit/index.html">ViT</a> attention
|
||||
and channel MLP is the <a href="../transformers/feed_forward.html">FFN</a> of ViT).</p>
|
||||
<p>ConvMixer use a $1 \times 1$ convolution for channel mixing and a
|
||||
depth-wise convolution for spatial mixing.
|
||||
Since it’s a convolution instead of a full MLP across the space, it mixes only the nearby batches in
|
||||
contrast to ViT or MLP-Mixer.
|
||||
Also the MLP-mixer uses MLPs of two layers for each mixing and ConvMixer uses a single layer for each mixing.</p>
|
||||
<p>The paper recommends removing the residual connection across the channel mixing (point-wise convolution),
|
||||
and having only a residual connection over the spatial mixing (depth-wise convolution).
|
||||
They also use <a href="../normalization/batch_norm/index.html">Batch normalization</a> instead
|
||||
of [Layer normalization)(../normalization/layer_norm/index.html).</p>
|
||||
<p>Here’s <a href="experiment.html">an experiment</a> that trains ConvMixer on CIFAR-10.</p>
|
||||
<p><a href="https://app.labml.ai/run/0fc344da2cd011ecb0bc3fdb2e774a3d"><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">38</span><span></span><span class="kn">import</span> <span class="nn">torch</span>
|
||||
<span class="lineno">39</span><span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">nn</span>
|
||||
<span class="lineno">40</span>
|
||||
<span class="lineno">41</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">42</span><span class="kn">from</span> <span class="nn">labml_nn.utils</span> <span class="kn">import</span> <span class="n">clone_module_list</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-1'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-1'>#</a>
|
||||
</div>
|
||||
<p><a id="ConvMixerLayer"></p>
|
||||
<h2>ConvMixer layer</h2>
|
||||
<p></a></p>
|
||||
<p>This is a single ConvMixer layer. The model will have a series of these.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">45</span><span class="k">class</span> <span class="nc">ConvMixerLayer</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 channels in patch embeddings, $h$</li>
|
||||
<li><code>kernel_size</code> is the size of the kernel of spatial convolution, $k$</li>
|
||||
</ul>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">54</span> <span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">d_model</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span> <span class="n">kernel_size</span><span class="p">:</span> <span class="nb">int</span><span class="p">):</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-3'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-3'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">59</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>Depth-wise convolution is separate convolution for each channel.
|
||||
We do this with a convolution layer with the number of groups equal to the number of channels.
|
||||
So that each channel is it’s own group.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">63</span> <span class="bp">self</span><span class="o">.</span><span class="n">depth_wise_conv</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Conv2d</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="lineno">64</span> <span class="n">kernel_size</span><span class="o">=</span><span class="n">kernel_size</span><span class="p">,</span>
|
||||
<span class="lineno">65</span> <span class="n">groups</span><span class="o">=</span><span class="n">d_model</span><span class="p">,</span>
|
||||
<span class="lineno">66</span> <span class="n">padding</span><span class="o">=</span><span class="p">(</span><span class="n">kernel_size</span> <span class="o">-</span> <span class="mi">1</span><span class="p">)</span> <span class="o">//</span> <span class="mi">2</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>Activation after depth-wise convolution</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">68</span> <span class="bp">self</span><span class="o">.</span><span class="n">act1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">GELU</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>Normalization after depth-wise convolution</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">70</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">BatchNorm2d</span><span class="p">(</span><span class="n">d_model</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>Point-wise convolution is a $1 \times 1$ convolution.
|
||||
i.e. a linear transformation of patch embeddings</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">74</span> <span class="bp">self</span><span class="o">.</span><span class="n">point_wise_conv</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Conv2d</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">kernel_size</span><span class="o">=</span><span class="mi">1</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>Activation after point-wise convolution</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">76</span> <span class="bp">self</span><span class="o">.</span><span class="n">act2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">GELU</span><span class="p">()</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-9'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-9'>#</a>
|
||||
</div>
|
||||
<p>Normalization after point-wise convolution</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">78</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">BatchNorm2d</span><span class="p">(</span><span class="n">d_model</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>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">80</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-11'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-11'>#</a>
|
||||
</div>
|
||||
<p>For the residual connection around the depth-wise convolution</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">82</span> <span class="n">residual</span> <span class="o">=</span> <span class="n">x</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>Depth-wise convolution, activation and normalization</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">85</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">depth_wise_conv</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
|
||||
<span class="lineno">86</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">act1</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
|
||||
<span class="lineno">87</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm1</span><span class="p">(</span><span class="n">x</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>Add residual connection</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">90</span> <span class="n">x</span> <span class="o">+=</span> <span class="n">residual</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>Point-wise convolution, activation and normalization</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">93</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">point_wise_conv</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
|
||||
<span class="lineno">94</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">act2</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
|
||||
<span class="lineno">95</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm2</span><span class="p">(</span><span class="n">x</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-15'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-15'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">98</span> <span class="k">return</span> <span class="n">x</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-16'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-16'>#</a>
|
||||
</div>
|
||||
<p><a id="PatchEmbeddings"></p>
|
||||
<h2>Get patch embeddings</h2>
|
||||
<p></a></p>
|
||||
<p>This splits the image into patches of size $p \times p$ and gives an embedding for each patch.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">101</span><span class="k">class</span> <span class="nc">PatchEmbeddings</span><span class="p">(</span><span class="n">Module</span><span class="p">):</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>
|
||||
<ul>
|
||||
<li><code>d_model</code> is the number of channels in patch embeddings $h$</li>
|
||||
<li><code>patch_size</code> is the size of the patch, $p$</li>
|
||||
<li><code>in_channels</code> is the number of channels in the input image (3 for rgb)</li>
|
||||
</ul>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">110</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">patch_size</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span> <span class="n">in_channels</span><span class="p">:</span> <span class="nb">int</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">116</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-19'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-19'>#</a>
|
||||
</div>
|
||||
<p>We create a convolution layer with a kernel size and and stride length equal to patch size.
|
||||
This is equivalent to splitting the image into patches and doing a linear
|
||||
transformation on each patch.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">121</span> <span class="bp">self</span><span class="o">.</span><span class="n">conv</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Conv2d</span><span class="p">(</span><span class="n">in_channels</span><span class="p">,</span> <span class="n">d_model</span><span class="p">,</span> <span class="n">kernel_size</span><span class="o">=</span><span class="n">patch_size</span><span class="p">,</span> <span class="n">stride</span><span class="o">=</span><span class="n">patch_size</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>Activation function</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">123</span> <span class="bp">self</span><span class="o">.</span><span class="n">act</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">GELU</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>Batch normalization</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">125</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">BatchNorm2d</span><span class="p">(</span><span class="n">d_model</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-22'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-22'>#</a>
|
||||
</div>
|
||||
<ul>
|
||||
<li><code>x</code> is the input image of shape <code>[batch_size, channels, height, width]</code></li>
|
||||
</ul>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">127</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-23'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-23'>#</a>
|
||||
</div>
|
||||
<p>Apply convolution layer</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">132</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">conv</span><span class="p">(</span><span class="n">x</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>Activation and normalization</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">134</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">act</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
|
||||
<span class="lineno">135</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm</span><span class="p">(</span><span class="n">x</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-25'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-25'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">138</span> <span class="k">return</span> <span class="n">x</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-26'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-26'>#</a>
|
||||
</div>
|
||||
<p><a id="ClassificationHead"></p>
|
||||
<h2>Classification Head</h2>
|
||||
<p></a></p>
|
||||
<p>They do average pooling (taking the mean of all patch embeddings) and a final linear transformation
|
||||
to predict the log-probabilities of the image classes.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">141</span><span class="k">class</span> <span class="nc">ClassificationHead</span><span class="p">(</span><span class="n">Module</span><span class="p">):</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-27'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-27'>#</a>
|
||||
</div>
|
||||
<ul>
|
||||
<li><code>d_model</code> is the number of channels in patch embeddings, $h$</li>
|
||||
<li><code>n_classes</code> is the number of classes in the classification task</li>
|
||||
</ul>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">151</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">n_classes</span><span class="p">:</span> <span class="nb">int</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>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">156</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-29'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-29'>#</a>
|
||||
</div>
|
||||
<p>Average Pool</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">158</span> <span class="bp">self</span><span class="o">.</span><span class="n">pool</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">AdaptiveAvgPool2d</span><span class="p">((</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">))</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-30'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-30'>#</a>
|
||||
</div>
|
||||
<p>Linear layer</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">160</span> <span class="bp">self</span><span class="o">.</span><span class="n">linear</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="n">d_model</span><span class="p">,</span> <span class="n">n_classes</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>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">162</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-32'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-32'>#</a>
|
||||
</div>
|
||||
<p>Average pooling</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">164</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">pool</span><span class="p">(</span><span class="n">x</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>Get the embedding, <code>x</code> will have shape <code>[batch_size, d_model, 1, 1]</code></p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">166</span> <span class="n">x</span> <span class="o">=</span> <span class="n">x</span><span class="p">[:,</span> <span class="p">:,</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-34'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-34'>#</a>
|
||||
</div>
|
||||
<p>Linear layer</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">168</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">linear</span><span class="p">(</span><span class="n">x</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-35'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-35'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">171</span> <span class="k">return</span> <span class="n">x</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-36'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-36'>#</a>
|
||||
</div>
|
||||
<h2>ConvMixer</h2>
|
||||
<p>This combines the patch embeddings block, a number of ConvMixer layers and a classification head.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">174</span><span class="k">class</span> <span class="nc">ConvMixer</span><span class="p">(</span><span class="n">Module</span><span class="p">):</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-37'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-37'>#</a>
|
||||
</div>
|
||||
<ul>
|
||||
<li><code>conv_mixer_layer</code> is a copy of a single <a href="#ConvMixerLayer">ConvMixer layer</a>.
|
||||
We make copies of it to make ConvMixer with <code>n_layers</code>.</li>
|
||||
<li><code>n_layers</code> is the number of ConvMixer layers (or depth), $d$.</li>
|
||||
<li><code>patch_emb</code> is the <a href="#PatchEmbeddings">patch embeddings layer</a>.</li>
|
||||
<li><code>classification</code> is the <a href="#ClassificationHead">classification head</a>.</li>
|
||||
</ul>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">181</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">conv_mixer_layer</span><span class="p">:</span> <span class="n">ConvMixerLayer</span><span class="p">,</span> <span class="n">n_layers</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
|
||||
<span class="lineno">182</span> <span class="n">patch_emb</span><span class="p">:</span> <span class="n">PatchEmbeddings</span><span class="p">,</span>
|
||||
<span class="lineno">183</span> <span class="n">classification</span><span class="p">:</span> <span class="n">ClassificationHead</span><span class="p">):</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-38'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-38'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">191</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-39'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-39'>#</a>
|
||||
</div>
|
||||
<p>Patch embeddings</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">193</span> <span class="bp">self</span><span class="o">.</span><span class="n">patch_emb</span> <span class="o">=</span> <span class="n">patch_emb</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-40'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-40'>#</a>
|
||||
</div>
|
||||
<p>Classification head</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">195</span> <span class="bp">self</span><span class="o">.</span><span class="n">classification</span> <span class="o">=</span> <span class="n">classification</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-41'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-41'>#</a>
|
||||
</div>
|
||||
<p>Make copies of the <a href="#ConvMixerLayer">ConvMixer layer</a></p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">197</span> <span class="bp">self</span><span class="o">.</span><span class="n">conv_mixer_layers</span> <span class="o">=</span> <span class="n">clone_module_list</span><span class="p">(</span><span class="n">conv_mixer_layer</span><span class="p">,</span> <span class="n">n_layers</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-42'>
|
||||
<div class='docs doc-strings'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-42'>#</a>
|
||||
</div>
|
||||
<ul>
|
||||
<li><code>x</code> is the input image of shape <code>[batch_size, channels, height, width]</code></li>
|
||||
</ul>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">199</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-43'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-43'>#</a>
|
||||
</div>
|
||||
<p>Get patch embeddings. This gives a tensor of shape <code>[batch_size, d_model, height / patch_size, width / patch_size]</code>.</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">204</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">patch_emb</span><span class="p">(</span><span class="n">x</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-44'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-44'>#</a>
|
||||
</div>
|
||||
<p>Pass through <a href="#ConvMixerLayer">ConvMixer layers</a></p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">207</span> <span class="k">for</span> <span class="n">layer</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">conv_mixer_layers</span><span class="p">:</span>
|
||||
<span class="lineno">208</span> <span class="n">x</span> <span class="o">=</span> <span class="n">layer</span><span class="p">(</span><span class="n">x</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-45'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-45'>#</a>
|
||||
</div>
|
||||
<p>Classification head, to get logits</p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">211</span> <span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">classification</span><span class="p">(</span><span class="n">x</span><span class="p">)</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='section' id='section-46'>
|
||||
<div class='docs'>
|
||||
<div class='section-link'>
|
||||
<a href='#section-46'>#</a>
|
||||
</div>
|
||||
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre><span class="lineno">214</span> <span class="k">return</span> <span class="n">x</span></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='footer'>
|
||||
<a href="https://papers.labml.ai">Trending Research Papers</a>
|
||||
<a href="https://labml.ai">labml.ai</a>
|
||||
</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>
|
||||
@ -599,18 +599,6 @@ q(x_t|x_0) &= \mathcal{N} \Big(x_t; \sqrt{\bar\alpha_t} x_0, (1-\bar\alpha_t) \m
|
||||
<div class="highlight"><pre><span class="lineno">287</span> <span class="k">return</span> <span class="n">F</span><span class="o">.</span><span class="n">mse_loss</span><span class="p">(</span><span class="n">noise</span><span class="p">,</span> <span class="n">eps_theta</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>
|
||||
<h2>Here’s our Twitter thread with a summary</h2>
|
||||
<p><blockquote class="twitter-tweet"><p lang="en" dir="ltr">Annotated <a href="https://twitter.com/PyTorch?ref_src=twsrc%5Etfw">@PyTorch</a> implementation of "Denoising Diffusion Probabilistic Models" by <a href="https://twitter.com/hojonathanho?ref_src=twsrc%5Etfw">@hojonathanho</a> <a href="https://twitter.com/ajayj_?ref_src=twsrc%5Etfw">@ajayj_</a> <a href="https://twitter.com/pabbeel?ref_src=twsrc%5Etfw">@pabbeel</a> <a href="https://twitter.com/berkeley_ai?ref_src=twsrc%5Etfw">@berkeley_ai</a><br><br>📝 Annotated code <a href="https://t.co/IxJMNQxJMa">https://t.co/IxJMNQxJMa</a><br>🖥 Github <a href="https://t.co/he5yIZZlB2">https://t.co/he5yIZZlB2</a><br>📎 Paper <a href="https://t.co/FjpamUVhLI">https://t.co/FjpamUVhLI</a><br><br>🧵👇 <a href="https://t.co/5SIZud6OnH">pic.twitter.com/5SIZud6OnH</a></p>— labml.ai (@labmlai) <a href="https://twitter.com/labmlai/status/1446676487361290240?ref_src=twsrc%5Etfw">October 9, 2021</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script></p>
|
||||
</div>
|
||||
<div class='code'>
|
||||
<div class="highlight"><pre></pre></div>
|
||||
</div>
|
||||
</div>
|
||||
<div class='footer'>
|
||||
<a href="https://papers.labml.ai">Trending Research Papers</a>
|
||||
<a href="https://labml.ai">labml.ai</a>
|
||||
|
||||
@ -102,6 +102,7 @@ implementations.</p>
|
||||
<h4>✨ <a href="lstm/index.html">LSTM</a></h4>
|
||||
<h4>✨ <a href="hypernetworks/hyper_lstm.html">HyperNetworks - HyperLSTM</a></h4>
|
||||
<h4>✨ <a href="resnet/index.html">ResNet</a></h4>
|
||||
<h4>✨ <a href="conv_mixer/index.html">ConvMixer</a></h4>
|
||||
<h4>✨ <a href="capsule_networks/index.html">Capsule Networks</a></h4>
|
||||
<h4>✨ <a href="gan/index.html">Generative Adversarial Networks</a></h4>
|
||||
<ul>
|
||||
|
||||
@ -344,7 +344,7 @@
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/diffusion/ddpm/index.html</loc>
|
||||
<lastmod>2021-10-08T16:30:00+00:00</lastmod>
|
||||
<lastmod>2021-10-09T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
@ -881,6 +881,20 @@
|
||||
</url>
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/conv_mixer/index.html</loc>
|
||||
<lastmod>2021-10-14T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/conv_mixer/experiment.html</loc>
|
||||
<lastmod>2021-10-14T16:30:00+00:00</lastmod>
|
||||
<priority>1.00</priority>
|
||||
</url>
|
||||
|
||||
|
||||
<url>
|
||||
<loc>https://nn.labml.ai/uncertainty/evidence/index.html</loc>
|
||||
<lastmod>2021-08-21T16:30:00+00:00</lastmod>
|
||||
|
||||
@ -42,6 +42,8 @@ implementations.
|
||||
|
||||
#### ✨ [ResNet](resnet/index.html)
|
||||
|
||||
#### ✨ [ConvMixer](conv_mixer/index.html)
|
||||
|
||||
#### ✨ [Capsule Networks](capsule_networks/index.html)
|
||||
|
||||
#### ✨ [Generative Adversarial Networks](gan/index.html)
|
||||
|
||||
214
labml_nn/conv_mixer/__init__.py
Normal file
214
labml_nn/conv_mixer/__init__.py
Normal file
@ -0,0 +1,214 @@
|
||||
"""
|
||||
---
|
||||
title: Patches Are All You Need? (ConvMixer)
|
||||
summary: >
|
||||
A PyTorch implementation/tutorial of the paper
|
||||
"Patches Are All You Need?"
|
||||
---
|
||||
|
||||
# Patches Are All You Need? (ConvMixer)
|
||||
|
||||
This is a [PyTorch](https://pytorch.org) implementation of the paper
|
||||
[Patches Are All You Need?](https://papers.labml.ai/paper/TVHS5Y4dNvM).
|
||||
|
||||
|
||||
|
||||
ConvMixer is Similar to [MLP-Mixer](../transformers/mlp_mixer/index.html).
|
||||
MLP-Mixer separates mixing of spatial and channel dimensions, by applying a MLP across spatial dimension
|
||||
and then an MLP across the channel dimension
|
||||
(spatial MLP replaces the [ViT](../transformers/vit/index.html) attention
|
||||
and channel MLP is the [FFN](../transformers/feed_forward.html) of ViT).
|
||||
|
||||
ConvMixer use a $1 \times 1$ convolution for channel mixing and a
|
||||
depth-wise convolution for spatial mixing.
|
||||
Since it's a convolution instead of a full MLP across the space, it mixes only the nearby batches in
|
||||
contrast to ViT or MLP-Mixer.
|
||||
Also the MLP-mixer uses MLPs of two layers for each mixing and ConvMixer uses a single layer for each mixing.
|
||||
|
||||
The paper recommends removing the residual connection across the channel mixing (point-wise convolution),
|
||||
and having only a residual connection over the spatial mixing (depth-wise convolution).
|
||||
They also use [Batch normalization](../normalization/batch_norm/index.html) instead
|
||||
of [Layer normalization)(../normalization/layer_norm/index.html).
|
||||
|
||||
Here's [an experiment](experiment.html) that trains ConvMixer on CIFAR-10.
|
||||
|
||||
[](https://app.labml.ai/run/0fc344da2cd011ecb0bc3fdb2e774a3d)
|
||||
"""
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
|
||||
from labml_helpers.module import Module
|
||||
from labml_nn.utils import clone_module_list
|
||||
|
||||
|
||||
class ConvMixerLayer(Module):
|
||||
"""
|
||||
<a id="ConvMixerLayer">
|
||||
## ConvMixer layer
|
||||
</a>
|
||||
|
||||
This is a single ConvMixer layer. The model will have a series of these.
|
||||
"""
|
||||
|
||||
def __init__(self, d_model: int, kernel_size: int):
|
||||
"""
|
||||
* `d_model` is the number of channels in patch embeddings, $h$
|
||||
* `kernel_size` is the size of the kernel of spatial convolution, $k$
|
||||
"""
|
||||
super().__init__()
|
||||
# Depth-wise convolution is separate convolution for each channel.
|
||||
# We do this with a convolution layer with the number of groups equal to the number of channels.
|
||||
# So that each channel is it's own group.
|
||||
self.depth_wise_conv = nn.Conv2d(d_model, d_model,
|
||||
kernel_size=kernel_size,
|
||||
groups=d_model,
|
||||
padding=(kernel_size - 1) // 2)
|
||||
# Activation after depth-wise convolution
|
||||
self.act1 = nn.GELU()
|
||||
# Normalization after depth-wise convolution
|
||||
self.norm1 = nn.BatchNorm2d(d_model)
|
||||
|
||||
# Point-wise convolution is a $1 \times 1$ convolution.
|
||||
# i.e. a linear transformation of patch embeddings
|
||||
self.point_wise_conv = nn.Conv2d(d_model, d_model, kernel_size=1)
|
||||
# Activation after point-wise convolution
|
||||
self.act2 = nn.GELU()
|
||||
# Normalization after point-wise convolution
|
||||
self.norm2 = nn.BatchNorm2d(d_model)
|
||||
|
||||
def forward(self, x: torch.Tensor):
|
||||
# For the residual connection around the depth-wise convolution
|
||||
residual = x
|
||||
|
||||
# Depth-wise convolution, activation and normalization
|
||||
x = self.depth_wise_conv(x)
|
||||
x = self.act1(x)
|
||||
x = self.norm1(x)
|
||||
|
||||
# Add residual connection
|
||||
x += residual
|
||||
|
||||
# Point-wise convolution, activation and normalization
|
||||
x = self.point_wise_conv(x)
|
||||
x = self.act2(x)
|
||||
x = self.norm2(x)
|
||||
|
||||
#
|
||||
return x
|
||||
|
||||
|
||||
class PatchEmbeddings(Module):
|
||||
"""
|
||||
<a id="PatchEmbeddings">
|
||||
## Get patch embeddings
|
||||
</a>
|
||||
|
||||
This splits the image into patches of size $p \times p$ and gives an embedding for each patch.
|
||||
"""
|
||||
|
||||
def __init__(self, d_model: int, patch_size: int, in_channels: int):
|
||||
"""
|
||||
* `d_model` is the number of channels in patch embeddings $h$
|
||||
* `patch_size` is the size of the patch, $p$
|
||||
* `in_channels` is the number of channels in the input image (3 for rgb)
|
||||
"""
|
||||
super().__init__()
|
||||
|
||||
# We create a convolution layer with a kernel size and and stride length equal to patch size.
|
||||
# This is equivalent to splitting the image into patches and doing a linear
|
||||
# transformation on each patch.
|
||||
self.conv = nn.Conv2d(in_channels, d_model, kernel_size=patch_size, stride=patch_size)
|
||||
# Activation function
|
||||
self.act = nn.GELU()
|
||||
# Batch normalization
|
||||
self.norm = nn.BatchNorm2d(d_model)
|
||||
|
||||
def forward(self, x: torch.Tensor):
|
||||
"""
|
||||
* `x` is the input image of shape `[batch_size, channels, height, width]`
|
||||
"""
|
||||
# Apply convolution layer
|
||||
x = self.conv(x)
|
||||
# Activation and normalization
|
||||
x = self.act(x)
|
||||
x = self.norm(x)
|
||||
|
||||
#
|
||||
return x
|
||||
|
||||
|
||||
class ClassificationHead(Module):
|
||||
"""
|
||||
<a id="ClassificationHead">
|
||||
## Classification Head
|
||||
</a>
|
||||
|
||||
They do average pooling (taking the mean of all patch embeddings) and a final linear transformation
|
||||
to predict the log-probabilities of the image classes.
|
||||
"""
|
||||
|
||||
def __init__(self, d_model: int, n_classes: int):
|
||||
"""
|
||||
* `d_model` is the number of channels in patch embeddings, $h$
|
||||
* `n_classes` is the number of classes in the classification task
|
||||
"""
|
||||
super().__init__()
|
||||
# Average Pool
|
||||
self.pool = nn.AdaptiveAvgPool2d((1, 1))
|
||||
# Linear layer
|
||||
self.linear = nn.Linear(d_model, n_classes)
|
||||
|
||||
def forward(self, x: torch.Tensor):
|
||||
# Average pooling
|
||||
x = self.pool(x)
|
||||
# Get the embedding, `x` will have shape `[batch_size, d_model, 1, 1]`
|
||||
x = x[:, :, 0, 0]
|
||||
# Linear layer
|
||||
x = self.linear(x)
|
||||
|
||||
#
|
||||
return x
|
||||
|
||||
|
||||
class ConvMixer(Module):
|
||||
"""
|
||||
## ConvMixer
|
||||
|
||||
This combines the patch embeddings block, a number of ConvMixer layers and a classification head.
|
||||
"""
|
||||
|
||||
def __init__(self, conv_mixer_layer: ConvMixerLayer, n_layers: int,
|
||||
patch_emb: PatchEmbeddings,
|
||||
classification: ClassificationHead):
|
||||
"""
|
||||
* `conv_mixer_layer` is a copy of a single [ConvMixer layer](#ConvMixerLayer).
|
||||
We make copies of it to make ConvMixer with `n_layers`.
|
||||
* `n_layers` is the number of ConvMixer layers (or depth), $d$.
|
||||
* `patch_emb` is the [patch embeddings layer](#PatchEmbeddings).
|
||||
* `classification` is the [classification head](#ClassificationHead).
|
||||
"""
|
||||
super().__init__()
|
||||
# Patch embeddings
|
||||
self.patch_emb = patch_emb
|
||||
# Classification head
|
||||
self.classification = classification
|
||||
# Make copies of the [ConvMixer layer](#ConvMixerLayer)
|
||||
self.conv_mixer_layers = clone_module_list(conv_mixer_layer, n_layers)
|
||||
|
||||
def forward(self, x: torch.Tensor):
|
||||
"""
|
||||
* `x` is the input image of shape `[batch_size, channels, height, width]`
|
||||
"""
|
||||
# Get patch embeddings. This gives a tensor of shape `[batch_size, d_model, height / patch_size, width / patch_size]`.
|
||||
x = self.patch_emb(x)
|
||||
|
||||
# Pass through [ConvMixer layers](#ConvMixerLayer)
|
||||
for layer in self.conv_mixer_layers:
|
||||
x = layer(x)
|
||||
|
||||
# Classification head, to get logits
|
||||
x = self.classification(x)
|
||||
|
||||
#
|
||||
return x
|
||||
87
labml_nn/conv_mixer/experiment.py
Normal file
87
labml_nn/conv_mixer/experiment.py
Normal file
@ -0,0 +1,87 @@
|
||||
"""
|
||||
---
|
||||
title: Train ConvMixer on CIFAR 10
|
||||
summary: >
|
||||
Train ConvMixer on CIFAR 10
|
||||
---
|
||||
|
||||
# Train a [ConvMixer](index.html) on CIFAR 10
|
||||
|
||||
This script trains a ConvMixer on CIFAR 10 dataset.
|
||||
|
||||
This is not an attempt to reproduce the results of the paper.
|
||||
The paper uses image augmentations
|
||||
present in [PyTorch Image Models (timm)](https://github.com/rwightman/pytorch-image-models)
|
||||
for training. We haven't done this for simplicity - which causes our validation accuracy to drop.
|
||||
|
||||
[](https://app.labml.ai/run/0fc344da2cd011ecb0bc3fdb2e774a3d)
|
||||
"""
|
||||
|
||||
from labml import experiment
|
||||
from labml.configs import option
|
||||
from labml_nn.experiments.cifar10 import CIFAR10Configs
|
||||
|
||||
|
||||
class Configs(CIFAR10Configs):
|
||||
"""
|
||||
## Configurations
|
||||
|
||||
We use [`CIFAR10Configs`](../experiments/cifar10.html) which defines all the
|
||||
dataset related configurations, optimizer, and a training loop.
|
||||
"""
|
||||
|
||||
# Size of a patch, $p$
|
||||
patch_size: int = 2
|
||||
# Number of channels in patch embeddings, $h$
|
||||
d_model: int = 256
|
||||
# Number of [ConvMixer layers](#ConvMixerLayer) or depth, $d$
|
||||
n_layers: int = 8
|
||||
# Kernel size of the depth-wise convolution, $k$
|
||||
kernel_size: int = 7
|
||||
# Number of classes in the task
|
||||
n_classes: int = 10
|
||||
|
||||
|
||||
@option(Configs.model)
|
||||
def _conv_mixer(c: Configs):
|
||||
"""
|
||||
### Create model
|
||||
"""
|
||||
from labml_nn.conv_mixer import ConvMixerLayer, ConvMixer, ClassificationHead, PatchEmbeddings
|
||||
|
||||
# Create ConvMixer
|
||||
return ConvMixer(ConvMixerLayer(c.d_model, c.kernel_size), c.n_layers,
|
||||
PatchEmbeddings(c.d_model, c.patch_size, 3),
|
||||
ClassificationHead(c.d_model, c.n_classes)).to(c.device)
|
||||
|
||||
|
||||
def main():
|
||||
# Create experiment
|
||||
experiment.create(name='ConvMixer', comment='cifar10')
|
||||
# Create configurations
|
||||
conf = Configs()
|
||||
# Load configurations
|
||||
experiment.configs(conf, {
|
||||
# Optimizer
|
||||
'optimizer.optimizer': 'Adam',
|
||||
'optimizer.learning_rate': 2.5e-4,
|
||||
|
||||
# Training epochs and batch size
|
||||
'epochs': 150,
|
||||
'train_batch_size': 64,
|
||||
|
||||
# Simple image augmentations
|
||||
'train_dataset': 'cifar10_train_augmented',
|
||||
# Do not augment images for validation
|
||||
'valid_dataset': 'cifar10_valid_no_augment',
|
||||
})
|
||||
# Set model for saving/loading
|
||||
experiment.add_pytorch_models({'model': conf.model})
|
||||
# Start the experiment and run the training loop
|
||||
with experiment.start():
|
||||
conf.run()
|
||||
|
||||
|
||||
#
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
@ -285,6 +285,3 @@ class DenoiseDiffusion:
|
||||
|
||||
# MSE loss
|
||||
return F.mse_loss(noise, eps_theta)
|
||||
|
||||
# ## Here's our Twitter thread with a summary
|
||||
# <blockquote class="twitter-tweet"><p lang="en" dir="ltr">Annotated <a href="https://twitter.com/PyTorch?ref_src=twsrc%5Etfw">@PyTorch</a> implementation of "Denoising Diffusion Probabilistic Models" by <a href="https://twitter.com/hojonathanho?ref_src=twsrc%5Etfw">@hojonathanho</a> <a href="https://twitter.com/ajayj_?ref_src=twsrc%5Etfw">@ajayj_</a> <a href="https://twitter.com/pabbeel?ref_src=twsrc%5Etfw">@pabbeel</a> <a href="https://twitter.com/berkeley_ai?ref_src=twsrc%5Etfw">@berkeley_ai</a><br><br>📝 Annotated code <a href="https://t.co/IxJMNQxJMa">https://t.co/IxJMNQxJMa</a><br>🖥 Github <a href="https://t.co/he5yIZZlB2">https://t.co/he5yIZZlB2</a><br>📎 Paper <a href="https://t.co/FjpamUVhLI">https://t.co/FjpamUVhLI</a><br><br>🧵👇 <a href="https://t.co/5SIZud6OnH">pic.twitter.com/5SIZud6OnH</a></p>— labml.ai (@labmlai) <a href="https://twitter.com/labmlai/status/1446676487361290240?ref_src=twsrc%5Etfw">October 9, 2021</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>
|
||||
|
||||
@ -47,6 +47,8 @@ implementations almost weekly.
|
||||
|
||||
#### ✨ [ResNet](https://nn.labml.ai/resnet/index.html)
|
||||
|
||||
#### ✨ [ConvMixer](https://nn.labml.ai/conv_mixer/index.html)
|
||||
|
||||
#### ✨ [Capsule Networks](https://nn.labml.ai/capsule_networks/index.html)
|
||||
|
||||
#### ✨ [Generative Adversarial Networks](https://nn.labml.ai/gan/index.html)
|
||||
|
||||
Reference in New Issue
Block a user