mirror of
https://github.com/labmlai/annotated_deep_learning_paper_implementations.git
synced 2025-10-31 10:48:49 +08:00
147 lines
6.3 KiB
HTML
147 lines
6.3 KiB
HTML
<!DOCTYPE html>
|
|
<html>
|
|
<head>
|
|
<meta http-equiv="content-type" content="text/html;charset=utf-8"/>
|
|
<meta name="viewport" content="width=device-width, initial-scale=1.0"/>
|
|
<meta name="description" content=""/>
|
|
|
|
<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="Compressive 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/compressive/readme.html"/>
|
|
<meta property="og:title" content="Compressive 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="Compressive Transformer"/>
|
|
<meta property="og:description" content=""/>
|
|
|
|
<title>Compressive Transformer</title>
|
|
<link rel="shortcut icon" href="/icon.png"/>
|
|
<link rel="stylesheet" href="../../pylit.css">
|
|
<link rel="canonical" href="https://nn.labml.ai/transformers/compressive/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">compressive</a>
|
|
</p>
|
|
<p>
|
|
|
|
<a href="https://github.com/lab-ml/labml_nn/tree/master/labml_nn/transformers/compressive/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/compressive/index.html">Compressive Transformer</a></h1>
|
|
<p>This is an implementation of
|
|
<a href="https://arxiv.org/abs/1911.05507">Compressive Transformers for Long-Range Sequence Modelling</a>
|
|
in <a href="https://pytorch.org">PyTorch</a>.</p>
|
|
<p>This is an extension of <a href="https://nn.labml.ai/transformers/xl/index.html">Transformer XL</a> where past memories
|
|
are compressed to give a longer attention range.
|
|
That is, the furthest $n_{cm} c$ memories are compressed into
|
|
$n_{cm}$ memories, where $c$ is the compression rate.</p>
|
|
<h2>Compression operation</h2>
|
|
<p>The compression operation is defined as
|
|
$f_c: \mathbb{R}^{nc \times d} \rightarrow \mathbb{R}^{n \times d}$.
|
|
The paper introduces multiple choices for $f_c$ and we have only implemented
|
|
1D convolution which seems to give the best results.
|
|
Each layer has a separate compression operation $f_c^{(i)}$ where
|
|
$i$ is the layer number.</p>
|
|
<h2>Training compression operation</h2>
|
|
<p>Since training compression with BPTT requires maintaining
|
|
a very large computational graph (many time steps), the paper proposes
|
|
an <em>auto-encoding loss</em> and an <em>attention reconstruction loss</em>.
|
|
The auto-encoding loss decodes the original memories from the compressed memories
|
|
and calculates the loss.
|
|
Attention reconstruction loss computes the multi-headed attention results
|
|
on the compressed memory and on uncompressed memory and gets a mean squared error
|
|
between them.
|
|
We have implemented the latter here since it gives better results.</p>
|
|
<p>This implementation uses pre-layer normalization
|
|
while the paper uses post-layer normalization.
|
|
Pre-layer norm does the layer norm before FFN[../feedforward.html) and
|
|
self-attention, and the pass-through in the residual connection is not normalized.
|
|
This is supposed to be more stable in standard transformer setups.</p>
|
|
<p>Here are <a href="https://nn.labml.ai/transformers/compressive/experiment.html">the training code</a> and a notebook for training a compressive transformer
|
|
model on the Tiny Shakespeare dataset.</p>
|
|
<p><a href="https://colab.research.google.com/github/lab-ml/nn/blob/master/labml_nn/transformers/compressive/experiment.ipynb"><img alt="Open In Colab" src="https://colab.research.google.com/assets/colab-badge.svg" /></a>
|
|
<a href="https://app.labml.ai/run/0d9b5338726c11ebb7c80242ac1c0002"><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>
|
|
</body>
|
|
</html> |