80 lines
8.4 KiB
HTML
80 lines
8.4 KiB
HTML
<!DOCTYPE html>
|
||
<html lang="en"><head><meta charset="UTF-8"/><meta name="viewport" content="width=device-width, initial-scale=1.0"/><title>Basics · Flux</title><script>(function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){
|
||
(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),
|
||
m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m)
|
||
})(window,document,'script','https://www.google-analytics.com/analytics.js','ga');
|
||
|
||
ga('create', 'UA-36890222-9', 'auto');
|
||
ga('send', 'pageview');
|
||
</script><link href="https://cdnjs.cloudflare.com/ajax/libs/normalize/4.2.0/normalize.min.css" rel="stylesheet" type="text/css"/><link href="https://fonts.googleapis.com/css?family=Lato|Roboto+Mono" rel="stylesheet" type="text/css"/><link href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/4.6.3/css/font-awesome.min.css" rel="stylesheet" type="text/css"/><link href="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/9.12.0/styles/default.min.css" rel="stylesheet" type="text/css"/><script>documenterBaseURL=".."</script><script src="https://cdnjs.cloudflare.com/ajax/libs/require.js/2.2.0/require.min.js" data-main="../assets/documenter.js"></script><script src="../siteinfo.js"></script><script src="../../versions.js"></script><link href="../assets/documenter.css" rel="stylesheet" type="text/css"/><link href="../../flux.css" rel="stylesheet" type="text/css"/></head><body><nav class="toc"><h1>Flux</h1><select id="version-selector" onChange="window.location.href=this.value" style="visibility: hidden"></select><form class="search" id="search-form" action="../search.html"><input id="search-query" name="q" type="text" placeholder="Search docs"/></form><ul><li><a class="toctext" href="../index.html">Home</a></li><li><span class="toctext">Building Models</span><ul><li class="current"><a class="toctext" href="basics.html">Basics</a><ul class="internal"><li><a class="toctext" href="#Taking-Gradients-1">Taking Gradients</a></li><li><a class="toctext" href="#Building-Layers-1">Building Layers</a></li><li><a class="toctext" href="#Stacking-It-Up-1">Stacking It Up</a></li></ul></li><li><a class="toctext" href="recurrence.html">Recurrence</a></li><li><a class="toctext" href="layers.html">Model Reference</a></li></ul></li><li><span class="toctext">Training Models</span><ul><li><a class="toctext" href="../training/optimisers.html">Optimisers</a></li><li><a class="toctext" href="../training/training.html">Training</a></li></ul></li><li><a class="toctext" href="../data/onehot.html">One-Hot Encoding</a></li><li><a class="toctext" href="../gpu.html">GPU Support</a></li><li><a class="toctext" href="../community.html">Community</a></li></ul></nav><article id="docs"><header><nav><ul><li>Building Models</li><li><a href="basics.html">Basics</a></li></ul><a class="edit-page" href="https://github.com/FluxML/Flux.jl/blob/master/docs/src/models/basics.md"><span class="fa"></span> Edit on GitHub</a></nav><hr/><div id="topbar"><span>Basics</span><a class="fa fa-bars" href="#"></a></div></header><h1><a class="nav-anchor" id="Model-Building-Basics-1" href="#Model-Building-Basics-1">Model-Building Basics</a></h1><h2><a class="nav-anchor" id="Taking-Gradients-1" href="#Taking-Gradients-1">Taking Gradients</a></h2><p>Consider a simple linear regression, which tries to predict an output array <code>y</code> from an input <code>x</code>. (It's a good idea to follow this example in the Julia repl.)</p><pre><code class="language-julia">W = rand(2, 5)
|
||
b = rand(2)
|
||
|
||
predict(x) = W*x .+ b
|
||
loss(x, y) = sum((predict(x) .- y).^2)
|
||
|
||
x, y = rand(5), rand(2) # Dummy data
|
||
loss(x, y) # ~ 3</code></pre><p>To improve the prediction we can take the gradients of <code>W</code> and <code>b</code> with respect to the loss function and perform gradient descent. We could calculate gradients by hand, but Flux will do it for us if we tell it that <code>W</code> and <code>b</code> are trainable <em>parameters</em>.</p><pre><code class="language-julia">using Flux.Tracker
|
||
|
||
W = param(W)
|
||
b = param(b)
|
||
|
||
l = loss(x, y)
|
||
|
||
back!(l)</code></pre><p><code>loss(x, y)</code> returns the same number, but it's now a <em>tracked</em> value that records gradients as it goes along. Calling <code>back!</code> then calculates the gradient of <code>W</code> and <code>b</code>. We can see what this gradient is, and modify <code>W</code> to train the model.</p><pre><code class="language-julia">W.grad
|
||
|
||
# Update the parameter
|
||
W.data .-= 0.1(W.grad)
|
||
|
||
loss(x, y) # ~ 2.5</code></pre><p>The loss has decreased a little, meaning that our prediction <code>x</code> is closer to the target <code>y</code>. If we have some data we can already try <a href="../training/training.html">training the model</a>.</p><p>All deep learning in Flux, however complex, is a simple generalisation of this example. Of course, models can <em>look</em> very different – they might have millions of parameters or complex control flow, and there are ways to manage this complexity. Let's see what that looks like.</p><h2><a class="nav-anchor" id="Building-Layers-1" href="#Building-Layers-1">Building Layers</a></h2><p>It's common to create more complex models than the linear regression above. For example, we might want to have two linear layers with a nonlinearity like <a href="https://en.wikipedia.org/wiki/Sigmoid_function">sigmoid</a> (<code>σ</code>) in between them. In the above style we could write this as:</p><pre><code class="language-julia">W1 = param(rand(3, 5))
|
||
b1 = param(rand(3))
|
||
layer1(x) = W1 * x .+ b1
|
||
|
||
W2 = param(rand(2, 3))
|
||
b2 = param(rand(2))
|
||
layer2(x) = W2 * x .+ b2
|
||
|
||
model(x) = layer2(σ.(layer1(x)))
|
||
|
||
model(rand(5)) # => 2-element vector</code></pre><p>This works but is fairly unwieldy, with a lot of repetition – especially as we add more layers. One way to factor this out is to create a function that returns linear layers.</p><pre><code class="language-julia">function linear(in, out)
|
||
W = param(randn(out, in))
|
||
b = param(randn(out))
|
||
x -> W * x .+ b
|
||
end
|
||
|
||
linear1 = linear(5, 3) # we can access linear1.W etc
|
||
linear2 = linear(3, 2)
|
||
|
||
model(x) = linear2(σ.(linear1(x)))
|
||
|
||
model(x) # => 2-element vector</code></pre><p>Another (equivalent) way is to create a struct that explicitly represents the affine layer.</p><pre><code class="language-julia">struct Affine
|
||
W
|
||
b
|
||
end
|
||
|
||
Affine(in::Integer, out::Integer) =
|
||
Affine(param(randn(out, in)), param(randn(out)))
|
||
|
||
# Overload call, so the object can be used as a function
|
||
(m::Affine)(x) = m.W * x .+ m.b
|
||
|
||
a = Affine(10, 5)
|
||
|
||
a(rand(10)) # => 5-element vector</code></pre><p>Congratulations! You just built the <code>Dense</code> layer that comes with Flux. Flux has many interesting layers available, but they're all things you could have built yourself very easily.</p><p>(There is one small difference with <code>Dense</code> – for convenience it also takes an activation function, like <code>Dense(10, 5, σ)</code>.)</p><h2><a class="nav-anchor" id="Stacking-It-Up-1" href="#Stacking-It-Up-1">Stacking It Up</a></h2><p>It's pretty common to write models that look something like:</p><pre><code class="language-julia">layer1 = Dense(10, 5, σ)
|
||
# ...
|
||
model(x) = layer3(layer2(layer1(x)))</code></pre><p>For long chains, it might be a bit more intuitive to have a list of layers, like this:</p><pre><code class="language-julia">using Flux
|
||
|
||
layers = [Dense(10, 5, σ), Dense(5, 2), softmax]
|
||
|
||
model(x) = foldl((x, m) -> m(x), x, layers)
|
||
|
||
model(rand(10)) # => 2-element vector</code></pre><p>Handily, this is also provided for in Flux:</p><pre><code class="language-julia">model2 = Chain(
|
||
Dense(10, 5, σ),
|
||
Dense(5, 2),
|
||
softmax)
|
||
|
||
model2(rand(10)) # => 2-element vector</code></pre><p>This quickly starts to look like a high-level deep learning library; yet you can see how it falls out of simple abstractions, and we lose none of the power of Julia code.</p><p>A nice property of this approach is that because "models" are just functions (possibly with trainable parameters), you can also see this as simple function composition.</p><pre><code class="language-julia">m = Dense(5, 2) ∘ Dense(10, 5, σ)
|
||
|
||
m(rand(10))</code></pre><p>Likewise, <code>Chain</code> will happily work with any Julia function.</p><pre><code class="language-julia">m = Chain(x -> x^2, x -> x+1)
|
||
|
||
m(5) # => 26</code></pre><footer><hr/><a class="previous" href="../index.html"><span class="direction">Previous</span><span class="title">Home</span></a><a class="next" href="recurrence.html"><span class="direction">Next</span><span class="title">Recurrence</span></a></footer></article></body></html>
|