<!DOCTYPE HTML>
<html lang="en">
<head>
<meta charset="UTF-8">
<title></title>
<meta content="text/html; charset=utf-8" http-equiv="Content-Type">
<meta name="description" content="">
<meta name="viewport" content="width=device-width, initial-scale=1">
<base href="">
<link rel="stylesheet" href="book.css">
<link href='http://fonts.googleapis.com/css?family=Open+Sans:300italic,400italic,600italic,700italic,800italic,400,300,600,700,800' rel='stylesheet' type='text/css'>
<link rel="shortcut icon" href="favicon.png">
<!-- Font Awesome -->
<link rel="stylesheet" href="http://maxcdn.bootstrapcdn.com/font-awesome/4.3.0/css/font-awesome.min.css">
<link rel="stylesheet" href="highlight.css">
<link rel="stylesheet" href="tomorrow-night.css">
<!-- MathJax -->
<script type="text/javascript" src="https://cdn.mathjax.org/mathjax/latest/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script>
<!-- Fetch JQuery from CDN but have a local fallback -->
<script src="http://code.jquery.com/jquery-2.1.4.min.js"></script>
<script>
if (typeof jQuery == 'undefined') {
document.write(unescape("%3Cscript src='jquery.js'%3E%3C/script%3E"));
}
</script>
</head>
<body>
<!-- Set the theme before any content is loaded, prevents flash -->
<script type="text/javascript">
var theme = localStorage.getItem('theme');
if (theme == null) { theme = 'light'; }
$('body').removeClass().addClass(theme);
</script>
<!-- Hide / unhide sidebar before it is displayed -->
<script type="text/javascript">
var sidebar = localStorage.getItem('sidebar');
if (sidebar === "hidden") { $("html").addClass("sidebar-hidden") }
else if (sidebar === "visible") { $("html").addClass("sidebar-visible") }
</script>
<div id="sidebar" class="sidebar">
<ul class="chapter"><li><a href="./juice.html"><strong>1.</strong> Juice</a></li><li><a href="./layers.html"><strong>2.</strong> Layers</a></li><li><ul class="section"><li><a href="./layer-lifecycle.html"><strong>2.1.</strong> Layer Lifecycle</a></li><li><a href="./building-networks.html"><strong>2.2.</strong> Create a Network</a></li><li><a href="./create-new-layer.html"><strong>2.3.</strong> Create a new Layer</a></li></ul></li><li><a href="./solvers.html"><strong>3.</strong> Solvers</a></li><li><ul class="section"><li><a href="./optimize-layers.html" class="active"><strong>3.1.</strong> Optimize Layers</a></li><li><a href="./multi-device-optimization.html"><strong>3.2.</strong> Multi-Device Optimization</a></li><li><a href="./distributed-optimization.html"><strong>3.3.</strong> Distributed Optimization</a></li></ul></li><li><a href="./backend.html"><strong>4.</strong> Backend</a></li><li><a href="./deep-learning-glossary.html"><strong>5.</strong> Glossary</a></li><li class="spacer"></li><li><a href="http://spearow.github.io/juice/juice/index.html"><strong>6.</strong> Rust API Documentation</a></li></ul>
</div>
<div id="page-wrapper" class="page-wrapper">
<div class="page">
<div id="menu-bar" class="menu-bar">
<div class="left-buttons">
<i id="sidebar-toggle" class="fa fa-bars"></i>
<i id="theme-toggle" class="fa fa-paint-brush"></i>
</div>
<h1 class="menu-title"></h1>
<div class="right-buttons">
<i id="print-button" class="fa fa-print" title="Print this book"></i>
</div>
</div>
<div id="content" class="content">
<h1>Optimize Layers</h1>
<p>In the previous chapter <a href="./solvers.html">3. Solver</a>, we learned what a solver
is and what it does. In this chapter, we take a look on how to optimize a
network via a <code>Solver</code>.</p>
<p>A <code>Solver</code> after its initialization has two <code>Layer</code>s, one for the <code>network</code>
which will be optimized and one for the <code>objective</code>. The output of the <code>network</code>
layer is used by the <code>objective</code> to compute the loss. The loss is then used by
the <code>Solver</code> to optimize the <code>network</code>.</p>
<p>The <code>Solver</code> has a very simple API - <code>.train_minibatch</code> and <code>.network</code>. The
optimization of the <code>network</code> is kicked off by the <code>.train_minibatch</code>
method, which takes two input parameters - some data that is feed to the network
and the expected target value for the network.</p>
<p>A SGD (Stochastic Gradient Descent) <code>Solver</code> would now compute the output of
the <code>network</code> using as input the data, put the output together with the expected
target value into the <code>objective</code> layer and use it, together with the gradient
of the <code>network</code> to optimize the weights of the <code>network</code>.</p>
<pre><code class="language-rust">/// Train the network with one minibatch
pub fn train_minibatch(&mut self, mb_data: ArcLock<SharedTensor<f32>>, mb_target: ArcLock<SharedTensor<f32>>) -> ArcLock<SharedTensor<f32>> {
// forward through network and classifier
let network_out = self.net.forward(&[mb_data])[0].clone();
let _ = self.objective.forward(&[network_out.clone(), mb_target]);
// forward through network and classifier
let classifier_gradient = self.objective.backward(&[]);
self.net.backward(&classifier_gradient[0 .. 1]);
self.worker.compute_update(&self.config, &mut self.net, self.iter);
self.net.update_weights(self.worker.backend());
self.iter += 1;
network_out
}
</code></pre>
<p>Using the <code>.train_minibatch</code> is straight forward. We pass the data as well as the
expected result of the <code>network</code> to the <code>.train_minibatch</code> method of the
initialized <code>Solver</code> struct. A more detailed example can be found at the
<a href="https://github.com/spearow/juice-examples">spearow/juice-examples</a> repository.</p>
<pre><code class="language-rust">let inp_lock = Arc::new(RwLock::new(inp));
let label_lock = Arc::new(RwLock::new(label));
// train the network!
let inferred_out = solver.train_minibatch(inp_lock.clone(), label_lock.clone());
</code></pre>
<p>If we don't want the <code>network</code> to be trained, we can use the <code>.network</code> method
of the <code>Solver</code> to receive access to the network. The <code>Solver</code> has actually
two network methods - <code>.network</code> and <code>mut_network</code>.</p>
<p>To run just the forward of the <code>network</code> without any optimization we can run</p>
<pre><code class="language-rust">let inferred_out = solver.network().forward(inp_lock.clone());
</code></pre>
<p>Juice ships with a <a href="https://en.wikipedia.org/wiki/Confusion_matrix">confusion matrix</a>,
which is a convenient way to visualize the performance of the optimized
<code>network</code>.</p>
<pre><code class="language-rust">let inferred_out = solver.train_minibatch(inp_lock.clone(), label_lock.clone());
let mut inferred = inferred_out.write().unwrap();
let predictions = confusion.get_predictions(&mut inferred);
confusion.add_samples(&predictions, &targets);
println!("Last sample: {} | Accuracy {}", confusion.samples().iter().last().unwrap(), confusion.accuracy());
</code></pre>
<p>A more detailed example can be found at the
<a href="https://github.com/spearow/juice-examples">spearow/juice-examples</a> repository.</p>
</div>
<!-- Mobile navigation buttons -->
<a href="./solvers.html" class="mobile-nav-chapters previous">
<i class="fa fa-angle-left"></i>
</a>
<a href="./multi-device-optimization.html" class="mobile-nav-chapters next">
<i class="fa fa-angle-right"></i>
</a>
</div>
<a href="./solvers.html" class="nav-chapters previous" title="You can navigate through the chapters using the arrow keys">
<i class="fa fa-angle-left"></i>
</a>
<a href="./multi-device-optimization.html" class="nav-chapters next" title="You can navigate through the chapters using the arrow keys">
<i class="fa fa-angle-right"></i>
</a>
</div>
<!-- Local fallback for Font Awesome -->
<script>
if ($(".fa").css("font-family") !== "FontAwesome") {
$('<link rel="stylesheet" type="text/css" href="_FontAwesome/css/font-awesome.css">').prependTo('head');
}
</script>
<script src="highlight.js"></script>
<script src="book.js"></script>
</body>
</html>
