wassname/pytorch-lightning
1
0
Fork 0
mirror of https://github.com/wassname/pytorch-lightning.git synced 2026-09-09 11:32:07 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
gh-pages
pytorch-lightning/Trainer/Checkpointing/index.html
T

756 lines
23 KiB
HTML
Raw Permalink Blame History

<!doctype html>
<html lang="en" class="no-js">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<meta http-equiv="x-ua-compatible" content="ie=edge">
<meta name="description" content="Documentation for PyTorch LightningModule, the researcher version of keras.">
<meta name="lang:clipboard.copy" content="Copy to clipboard">
<meta name="lang:clipboard.copied" content="Copied to clipboard">
<meta name="lang:search.language" content="en">
<meta name="lang:search.pipeline.stopwords" content="True">
<meta name="lang:search.pipeline.trimmer" content="True">
<meta name="lang:search.result.none" content="No matching documents">
<meta name="lang:search.result.one" content="1 matching document">
<meta name="lang:search.result.other" content="# matching documents">
<meta name="lang:search.tokenizer" content="[\s\-]+">
<link rel="shortcut icon" href="../../assets/images/favicon.png">
<meta name="generator" content="mkdocs-1.0.4, mkdocs-material-4.4.0">
<title>Checkpointing - PyTorch lightning Documentation</title>
<link rel="stylesheet" href="../../assets/stylesheets/application.0284f74d.css">
<script src="../../assets/javascripts/modernizr.74668098.js"></script>
<link href="https://fonts.gstatic.com" rel="preconnect" crossorigin>
<link rel="stylesheet" href="https://fonts.googleapis.com/css?family=Roboto:300,400,400i,700|Roboto+Mono&display=fallback">
<style>body,input{font-family:"Roboto","Helvetica Neue",Helvetica,Arial,sans-serif}code,kbd,pre{font-family:"Roboto Mono","Courier New",Courier,monospace}</style>
<link rel="stylesheet" href="../../assets/fonts/material-icons.css">
</head>
<body dir="ltr">
<svg class="md-svg">
<defs>
<svg xmlns="http://www.w3.org/2000/svg" width="416" height="448" viewBox="0 0 416 448" id="__github"><path fill="currentColor" d="M160 304q0 10-3.125 20.5t-10.75 19T128 352t-18.125-8.5-10.75-19T96 304t3.125-20.5 10.75-19T128 256t18.125 8.5 10.75 19T160 304zm160 0q0 10-3.125 20.5t-10.75 19T288 352t-18.125-8.5-10.75-19T256 304t3.125-20.5 10.75-19T288 256t18.125 8.5 10.75 19T320 304zm40 0q0-30-17.25-51T296 232q-10.25 0-48.75 5.25Q229.5 240 208 240t-39.25-2.75Q130.75 232 120 232q-29.5 0-46.75 21T56 304q0 22 8 38.375t20.25 25.75 30.5 15 35 7.375 37.25 1.75h42q20.5 0 37.25-1.75t35-7.375 30.5-15 20.25-25.75T360 304zm56-44q0 51.75-15.25 82.75-9.5 19.25-26.375 33.25t-35.25 21.5-42.5 11.875-42.875 5.5T212 416q-19.5 0-35.5-.75t-36.875-3.125-38.125-7.5-34.25-12.875T37 371.5t-21.5-28.75Q0 312 0 260q0-59.25 34-99-6.75-20.5-6.75-42.5 0-29 12.75-54.5 27 0 47.5 9.875t47.25 30.875Q171.5 96 212 96q37 0 70 8 26.25-20.5 46.75-30.25T376 64q12.75 25.5 12.75 54.5 0 21.75-6.75 42 34 40 34 99.5z"/></svg>
</defs>
</svg>
<input class="md-toggle" data-md-toggle="drawer" type="checkbox" id="__drawer" autocomplete="off">
<input class="md-toggle" data-md-toggle="search" type="checkbox" id="__search" autocomplete="off">
<label class="md-overlay" data-md-component="overlay" for="__drawer"></label>
<a href="#model-saving" tabindex="1" class="md-skip">
Skip to content
</a>
<header class="md-header" data-md-component="header">
<nav class="md-header-nav md-grid">
<div class="md-flex">
<div class="md-flex__cell md-flex__cell--shrink">
<a href="../.." title="PyTorch lightning Documentation" class="md-header-nav__button md-logo">
<i class="md-icon"></i>
</a>
</div>
<div class="md-flex__cell md-flex__cell--shrink">
<label class="md-icon md-icon--menu md-header-nav__button" for="__drawer"></label>
</div>
<div class="md-flex__cell md-flex__cell--stretch">
<div class="md-flex__ellipsis md-header-nav__title" data-md-component="title">
<span class="md-header-nav__topic">
PyTorch lightning Documentation
</span>
<span class="md-header-nav__topic">
Checkpointing
</span>
</div>
</div>
<div class="md-flex__cell md-flex__cell--shrink">
<label class="md-icon md-icon--search md-header-nav__button" for="__search"></label>
<div class="md-search" data-md-component="search" role="dialog">
<label class="md-search__overlay" for="__search"></label>
<div class="md-search__inner" role="search">
<form class="md-search__form" name="search">
<input type="text" class="md-search__input" name="query" placeholder="Search" autocapitalize="off" autocorrect="off" autocomplete="off" spellcheck="false" data-md-component="query" data-md-state="active">
<label class="md-icon md-search__icon" for="__search"></label>
<button type="reset" class="md-icon md-search__icon" data-md-component="reset" tabindex="-1">
&#xE5CD;
</button>
</form>
<div class="md-search__output">
<div class="md-search__scrollwrap" data-md-scrollfix>
<div class="md-search-result" data-md-component="result">
<div class="md-search-result__meta">
Type to start searching
</div>
<ol class="md-search-result__list"></ol>
</div>
</div>
</div>
</div>
</div>
</div>
<div class="md-flex__cell md-flex__cell--shrink">
<div class="md-header-nav__source">
<a href="https://github.com/williamFalcon/pytorch-lightning/" title="Go to repository" class="md-source" data-md-source="github">
<div class="md-source__icon">
<svg viewBox="0 0 24 24" width="24" height="24">
<use xlink:href="#__github" width="24" height="24"></use>
</svg>
</div>
<div class="md-source__repository">
williamFalcon/pytorch-lightning
</div>
</a>
</div>
</div>
</div>
</nav>
</header>
<div class="md-container">
<main class="md-main">
<div class="md-main__inner md-grid" data-md-component="container">
<div class="md-sidebar md-sidebar--primary" data-md-component="navigation">
<div class="md-sidebar__scrollwrap">
<div class="md-sidebar__inner">
<nav class="md-nav md-nav--primary" data-md-level="0">
<label class="md-nav__title md-nav__title--site" for="__drawer">
<a href="../.." title="PyTorch lightning Documentation" class="md-nav__button md-logo">
<i class="md-icon"></i>
</a>
PyTorch lightning Documentation
</label>
<div class="md-nav__source">
<a href="https://github.com/williamFalcon/pytorch-lightning/" title="Go to repository" class="md-source" data-md-source="github">
<div class="md-source__icon">
<svg viewBox="0 0 24 24" width="24" height="24">
<use xlink:href="#__github" width="24" height="24"></use>
</svg>
</div>
<div class="md-source__repository">
williamFalcon/pytorch-lightning
</div>
</a>
</div>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="../.." title="Home" class="md-nav__link">
Home
</a>
</li>
<li class="md-nav__item md-nav__item--nested">
<input class="md-toggle md-nav__toggle" data-md-toggle="nav-2" type="checkbox" id="nav-2">
<label class="md-nav__link" for="nav-2">
LightningModule
</label>
<nav class="md-nav" data-md-component="collapsible" data-md-level="1">
<label class="md-nav__title" for="nav-2">
LightningModule
</label>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="../../LightningModule/RequiredTrainerInterface/" title="Lightning Module interface" class="md-nav__link">
Lightning Module interface
</a>
</li>
<li class="md-nav__item">
<a href="../../LightningModule/methods/" title="Methods" class="md-nav__link">
Methods
</a>
</li>
<li class="md-nav__item">
<a href="../../LightningModule/properties/" title="Properties" class="md-nav__link">
Properties
</a>
</li>
</ul>
</nav>
</li>
<li class="md-nav__item md-nav__item--active md-nav__item--nested">
<input class="md-toggle md-nav__toggle" data-md-toggle="nav-3" type="checkbox" id="nav-3" checked>
<label class="md-nav__link" for="nav-3">
Trainer
</label>
<nav class="md-nav" data-md-component="collapsible" data-md-level="1">
<label class="md-nav__title" for="nav-3">
Trainer
</label>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="../" title="Trainer" class="md-nav__link">
Trainer
</a>
</li>
<li class="md-nav__item md-nav__item--active">
<input class="md-toggle md-nav__toggle" data-md-toggle="toc" type="checkbox" id="__toc">
<label class="md-nav__link md-nav__link--active" for="__toc">
Checkpointing
</label>
<a href="./" title="Checkpointing" class="md-nav__link md-nav__link--active">
Checkpointing
</a>
<nav class="md-nav md-nav--secondary">
<label class="md-nav__title" for="__toc">Table of contents</label>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="#model-saving" title="Model saving" class="md-nav__link">
Model saving
</a>
</li>
<li class="md-nav__item">
<a href="#restoring-training-session" title="Restoring training session" class="md-nav__link">
Restoring training session
</a>
</li>
</ul>
</nav>
</li>
<li class="md-nav__item">
<a href="../Distributed training/" title="Distributed training" class="md-nav__link">
Distributed training
</a>
</li>
<li class="md-nav__item">
<a href="../Logging/" title="Logging" class="md-nav__link">
Logging
</a>
</li>
<li class="md-nav__item">
<a href="../SLURM Managed Cluster/" title="SLURM Managed Cluster" class="md-nav__link">
SLURM Managed Cluster
</a>
</li>
<li class="md-nav__item">
<a href="../Testing loop/" title="Testing loop" class="md-nav__link">
Testing loop
</a>
</li>
<li class="md-nav__item">
<a href="../Training Loop/" title="Training Loop" class="md-nav__link">
Training Loop
</a>
</li>
<li class="md-nav__item">
<a href="../Validation loop/" title="Validation loop" class="md-nav__link">
Validation loop
</a>
</li>
<li class="md-nav__item">
<a href="../debugging/" title="Debugging" class="md-nav__link">
Debugging
</a>
</li>
<li class="md-nav__item">
<a href="../hooks/" title="Hooks" class="md-nav__link">
Hooks
</a>
</li>
</ul>
</nav>
</li>
<li class="md-nav__item md-nav__item--nested">
<input class="md-toggle md-nav__toggle" data-md-toggle="nav-4" type="checkbox" id="nav-4">
<label class="md-nav__link" for="nav-4">
Examples
</label>
<nav class="md-nav" data-md-component="collapsible" data-md-level="1">
<label class="md-nav__title" for="nav-4">
Examples
</label>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="../../examples/Examples/" title="Examples" class="md-nav__link">
Examples
</a>
</li>
</ul>
</nav>
</li>
</ul>
</nav>
</div>
</div>
</div>
<div class="md-sidebar md-sidebar--secondary" data-md-component="toc">
<div class="md-sidebar__scrollwrap">
<div class="md-sidebar__inner">
<nav class="md-nav md-nav--secondary">
<label class="md-nav__title" for="__toc">Table of contents</label>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="#model-saving" title="Model saving" class="md-nav__link">
Model saving
</a>
</li>
<li class="md-nav__item">
<a href="#restoring-training-session" title="Restoring training session" class="md-nav__link">
Restoring training session
</a>
</li>
</ul>
</nav>
</div>
</div>
</div>
<div class="md-content">
<article class="md-content__inner md-typeset">
<a href="https://github.com/williamFalcon/pytorch-lightning/edit/master/docs/Trainer/Checkpointing.md" title="Edit this page" class="md-icon md-content__icon">&#xE3C9;</a>
<h1>Checkpointing</h1>
<p>Lightning can automate saving and loading checkpoints.</p>
<hr />
<h3 id="model-saving">Model saving</h3>
<p>Checkpointing is enabled by default to the current working directory.
To change the checkpoint path pass in :</p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre>1</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="n">Trainer</span><span class="p">(</span><span class="n">default_save_path</span><span class="o">=</span><span class="s1">&#39;/your/path/to/save/checkpoints&#39;</span><span class="p">)</span>
</pre></div>
</td></tr></table>
<p>To modify the behavior of checkpointing pass in your own callback.</p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre> 1
2
3
4
5
6
7
8
9
10
11
12
13</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="kn">from</span> <span class="nn">pytorch_lightning.callbacks</span> <span class="kn">import</span> <span class="n">ModelCheckpoint</span>
<span class="c1"># DEFAULTS used by the Trainer</span>
<span class="n">checkpoint_callback</span> <span class="o">=</span> <span class="n">ModelCheckpoint</span><span class="p">(</span>
<span class="n">filepath</span><span class="o">=</span><span class="n">os</span><span class="o">.</span><span class="n">getcwd</span><span class="p">(),</span>
<span class="n">save_best_only</span><span class="o">=</span><span class="bp">True</span><span class="p">,</span>
<span class="n">verbose</span><span class="o">=</span><span class="bp">True</span><span class="p">,</span>
<span class="n">monitor</span><span class="o">=</span><span class="s1">&#39;val_loss&#39;</span><span class="p">,</span>
<span class="n">mode</span><span class="o">=</span><span class="s1">&#39;min&#39;</span><span class="p">,</span>
<span class="n">prefix</span><span class="o">=</span><span class="s1">&#39;&#39;</span>
<span class="p">)</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span><span class="n">checkpoint_callback</span><span class="o">=</span><span class="n">checkpoint_callback</span><span class="p">)</span>
</pre></div>
</td></tr></table>
<hr />
<h3 id="restoring-training-session">Restoring training session</h3>
<p>You might want to not only load a model but also continue training it. Use this method to
restore the trainer state as well. This will continue from the epoch and global step you last left off.<br />
However, the dataloaders will start from the first batch again (if you shuffled it shouldn't matter). </p>
<p>Lightning will restore the session if you pass a logger with the same version and there's a saved checkpoint. </p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre> 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="kn">from</span> <span class="nn">pytorch_lightning</span> <span class="kn">import</span> <span class="n">Trainer</span>
<span class="kn">from</span> <span class="nn">pytorch_lightning.logging</span> <span class="kn">import</span> <span class="n">TestTubeLogger</span>
<span class="n">logger</span> <span class="o">=</span> <span class="n">TestTubeLogger</span><span class="p">(</span>
<span class="n">save_dir</span><span class="o">=</span><span class="s1">&#39;./savepath&#39;</span><span class="p">,</span>
<span class="n">version</span><span class="o">=</span><span class="mi">1</span> <span class="c1"># An existing version with a saved checkpoint</span>
<span class="p">)</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span>
<span class="n">logger</span><span class="o">=</span><span class="n">logger</span><span class="p">,</span>
<span class="n">default_save_path</span><span class="o">=</span><span class="s1">&#39;./savepath&#39;</span>
<span class="p">)</span>
<span class="c1"># this fit call loads model weights and trainer state</span>
<span class="c1"># the trainer continues seamlessly from where you left off</span>
<span class="c1"># without having to do anything else.</span>
<span class="n">trainer</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
</pre></div>
</td></tr></table>
<p>The trainer restores: </p>
<ul>
<li>global_step </li>
<li>current_epoch </li>
<li>All optimizers </li>
<li>All lr_schedulers </li>
<li>Model weights</li>
</ul>
<p>You can even change the logic of your model as long as the weights and "architecture" of
the system isn't different. If you add a layer, for instance, it might not work. </p>
<p>At a rough level, here's <a href="https://github.com/williamFalcon/pytorch-lightning/blob/master/pytorch_lightning/root_module/model_saving.py#L63">what happens inside Trainer</a>: </p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre> 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="bp">self</span><span class="o">.</span><span class="n">global_step</span> <span class="o">=</span> <span class="n">checkpoint</span><span class="p">[</span><span class="s1">&#39;global_step&#39;</span><span class="p">]</span>
<span class="bp">self</span><span class="o">.</span><span class="n">current_epoch</span> <span class="o">=</span> <span class="n">checkpoint</span><span class="p">[</span><span class="s1">&#39;epoch&#39;</span><span class="p">]</span>
<span class="c1"># restore the optimizers</span>
<span class="n">optimizer_states</span> <span class="o">=</span> <span class="n">checkpoint</span><span class="p">[</span><span class="s1">&#39;optimizer_states&#39;</span><span class="p">]</span>
<span class="k">for</span> <span class="n">optimizer</span><span class="p">,</span> <span class="n">opt_state</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">optimizers</span><span class="p">,</span> <span class="n">optimizer_states</span><span class="p">):</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">opt_state</span><span class="p">)</span>
<span class="c1"># restore the lr schedulers</span>
<span class="n">lr_schedulers</span> <span class="o">=</span> <span class="n">checkpoint</span><span class="p">[</span><span class="s1">&#39;lr_schedulers&#39;</span><span class="p">]</span>
<span class="k">for</span> <span class="n">scheduler</span><span class="p">,</span> <span class="n">lrs_state</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">lr_schedulers</span><span class="p">,</span> <span class="n">lr_schedulers</span><span class="p">):</span>
<span class="n">scheduler</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">lrs_state</span><span class="p">)</span>
<span class="c1"># uses the model you passed into trainer </span>
<span class="n">model</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">checkpoint</span><span class="p">[</span><span class="s1">&#39;state_dict&#39;</span><span class="p">])</span>
</pre></div>
</td></tr></table>
</article>
</div>
</div>
</main>
<footer class="md-footer">
<div class="md-footer-nav">
<nav class="md-footer-nav__inner md-grid">
<a href="../" title="Trainer" class="md-flex md-footer-nav__link md-footer-nav__link--prev" rel="prev">
<div class="md-flex__cell md-flex__cell--shrink">
<i class="md-icon md-icon--arrow-back md-footer-nav__button"></i>
</div>
<div class="md-flex__cell md-flex__cell--stretch md-footer-nav__title">
<span class="md-flex__ellipsis">
<span class="md-footer-nav__direction">
Previous
</span>
Trainer
</span>
</div>
</a>
<a href="../Distributed training/" title="Distributed training" class="md-flex md-footer-nav__link md-footer-nav__link--next" rel="next">
<div class="md-flex__cell md-flex__cell--stretch md-footer-nav__title">
<span class="md-flex__ellipsis">
<span class="md-footer-nav__direction">
Next
</span>
Distributed training
</span>
</div>
<div class="md-flex__cell md-flex__cell--shrink">
<i class="md-icon md-icon--arrow-forward md-footer-nav__button"></i>
</div>
</a>
</nav>
</div>
<div class="md-footer-meta md-typeset">
<div class="md-footer-meta__inner md-grid">
<div class="md-footer-copyright">
powered by
<a href="https://www.mkdocs.org">MkDocs</a>
and
<a href="https://squidfunk.github.io/mkdocs-material/">
Material for MkDocs</a>
</div>
</div>
</div>
</footer>
</div>
<script src="../../assets/javascripts/application.245445c6.js"></script>
<script>app.initialize({version:"1.0.4",url:{base:"../.."}})</script>
</body>
</html>
Reference in New Issue View Git Blame Copy Permalink