build based on 3285afa
This commit is contained in:
parent
a23fd0f6dc
commit
17871f4723
@ -45,11 +45,18 @@ julia> W.grad
|
||||
julia> x.grad
|
||||
2-element Array{Float64,1}:
|
||||
-2.0
|
||||
-2.0</code></pre><p>You may sometimes want to drop derivative information and just get the plain value back. You can do this by calling <code>Tracker.data(W)</code>.</p><h2><a class="nav-anchor" id="Custom-Gradients-1" href="#Custom-Gradients-1">Custom Gradients</a></h2><p>We can hook in to the processes above to implement custom gradients for a function or kernel. For a toy example, imagine a custom implementation of <code>minus</code>:</p><pre><code class="language-julia">minus(a, b) = a - b</code></pre><p>Firstly, we must tell the tracker system to stop when it sees a call to <code>minus</code>, and record it. We can do this using dispatch:</p><pre><code class="language-julia">using Flux.Tracker: TrackedReal, track, @grad
|
||||
-2.0</code></pre><p>You may sometimes want to drop derivative information and just get the plain value back. You can do this by calling <code>Tracker.data(W)</code>.</p><h2><a class="nav-anchor" id="Custom-Gradients-1" href="#Custom-Gradients-1">Custom Gradients</a></h2><p>We can hook in to the processes above to implement custom gradients for a function or kernel. For a toy example, imagine a custom implementation of <code>minus</code>:</p><pre><code class="language-julia">minus(a, b) = a - b</code></pre><p>Firstly, we must tell the tracker system to stop when it sees a call to <code>minus</code>, and record it. We can do this using dispatch:</p><pre><code class="language-julia">using Flux.Tracker: TrackedArray, track, @grad
|
||||
|
||||
minus(a::TrackedArray, b::TrackedArray) = Tracker.track(minus, a, b)</code></pre><p><code>track</code> takes care of building a new <code>Tracked</code> object and recording the operation on the tape. We just need to provide a gradient definition.</p><pre><code class="language-julia">@grad function minus(a, b)
|
||||
return minus(data(a),data(b)), Δ -> (Δ, -Δ)
|
||||
end</code></pre><p>This is essentially just a way of overloading the <code>forward</code> function we saw above. We strip tracking from <code>a</code> and <code>b</code> so that we are calling the original definition of <code>minus</code> (otherwise, we'd just try to track the call again and hit an infinite regress).</p><p>Note that in the backpropagator we don't call <code>data(a)</code>; we <em>do</em> in fact want to track this, since nest AD will take a derivative through the backpropagator itself. For example, the gradient of <code>*</code> might look like this.</p><pre><code class="language-julia">@grad a * b = data(a)*data(b), Δ -> (Δ*b, a*Δ)</code></pre><p>For multi-argument functions with custom gradients, you likely want to catch not just <code>minus(::TrackedArray, ::TrackedArray)</code> but also <code>minus(::Array, TrackedArray)</code> and so on. To do so, just define those extra signatures as needed:</p><pre><code class="language-julia">minus(a::AbstractArray, b::TrackedArray) = Tracker.track(minus, a, b)
|
||||
minus(a::TrackedArray, b::TrackedArray) = track(minus, a, b)</code></pre><p><code>track</code> takes care of building a new <code>Tracked</code> object and recording the operation on the tape. We just need to provide a gradient definition.</p><pre><code class="language-julia">@grad function minus(a, b)
|
||||
return minus(data(a), data(b)), Δ -> (Δ, -Δ)
|
||||
end</code></pre><p>This is essentially just a way of overloading the <code>forward</code> function we saw above. We strip tracking from <code>a</code> and <code>b</code> so that we are calling the original definition of <code>minus</code> (otherwise, we'd just try to track the call again and hit an infinite regress).</p><p>Note that in the backpropagator we don't call <code>data(a)</code>; we <em>do</em> in fact want to track this, since nest AD will take a derivative through the backpropagator itself. For example, the gradient of <code>*</code> might look like this.</p><pre><code class="language-julia">@grad a * b = data(a)*data(b), Δ -> (Δ*b, a*Δ)</code></pre><p>We can then calculate the first derivative of <code>minus</code> as follows:</p><pre><code class="language-julia">a = param([1,2,3])
|
||||
b = param([3,2,1])
|
||||
|
||||
c = minus(a, b) # [-2.0 (tracked), 0.0 (tracked), 2.0 (tracked)]
|
||||
|
||||
Tracker.back!(c, 1)
|
||||
Tracker.grad(a) # [1.00, 1.00, 1.00]
|
||||
Tracker.grad(b) # [-1.00, -1.00, -1.00]</code></pre><p>For multi-argument functions with custom gradients, you likely want to catch not just <code>minus(::TrackedArray, ::TrackedArray)</code> but also <code>minus(::Array, TrackedArray)</code> and so on. To do so, just define those extra signatures as needed:</p><pre><code class="language-julia">minus(a::AbstractArray, b::TrackedArray) = Tracker.track(minus, a, b)
|
||||
minus(a::TrackedArray, b::AbstractArray) = Tracker.track(minus, a, b)</code></pre><h2><a class="nav-anchor" id="Tracked-Internals-1" href="#Tracked-Internals-1">Tracked Internals</a></h2><p>All <code>Tracked*</code> objects (<code>TrackedArray</code>, <code>TrackedReal</code>) are light wrappers around the <code>Tracked</code> type, which you can access via the <code>.tracker</code> field.</p><pre><code class="language-julia">julia> x.tracker
|
||||
Flux.Tracker.Tracked{Array{Float64,1}}(0x00000000, Flux.Tracker.Call{Nothing,Tuple{}}(nothing, ()), true, [5.0, 6.0], [-2.0, -2.0])</code></pre><p>The <code>Tracker</code> stores the gradient of a given object, which we've seen before.</p><pre><code class="language-julia">julia> x.tracker.grad
|
||||
2-element Array{Float64,1}:
|
||||
|
@ -11,26 +11,26 @@ m(5) == 26
|
||||
|
||||
m = Chain(Dense(10, 5), Dense(5, 2))
|
||||
x = rand(10)
|
||||
m(x) == m[2](m[1](x))</code></pre><p><code>Chain</code> also supports indexing and slicing, e.g. <code>m[2]</code> or <code>m[1:end-1]</code>. <code>m[1:3](x)</code> will calculate the output of the first three layers.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/basic.jl#L1-L18">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Dense" href="#Flux.Dense"><code>Flux.Dense</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">Dense(in::Integer, out::Integer, σ = identity)</code></pre><p>Creates a traditional <code>Dense</code> layer with parameters <code>W</code> and <code>b</code>.</p><pre><code class="language-none">y = σ.(W * x .+ b)</code></pre><p>The input <code>x</code> must be a vector of length <code>in</code>, or a batch of vectors represented as an <code>in × N</code> matrix. The out <code>y</code> will be a vector or batch of length <code>out</code>.</p><pre><code class="language-julia">julia> d = Dense(5, 2)
|
||||
m(x) == m[2](m[1](x))</code></pre><p><code>Chain</code> also supports indexing and slicing, e.g. <code>m[2]</code> or <code>m[1:end-1]</code>. <code>m[1:3](x)</code> will calculate the output of the first three layers.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/basic.jl#L1-L18">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Dense" href="#Flux.Dense"><code>Flux.Dense</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">Dense(in::Integer, out::Integer, σ = identity)</code></pre><p>Creates a traditional <code>Dense</code> layer with parameters <code>W</code> and <code>b</code>.</p><pre><code class="language-none">y = σ.(W * x .+ b)</code></pre><p>The input <code>x</code> must be a vector of length <code>in</code>, or a batch of vectors represented as an <code>in × N</code> matrix. The out <code>y</code> will be a vector or batch of length <code>out</code>.</p><pre><code class="language-julia">julia> d = Dense(5, 2)
|
||||
Dense(5, 2)
|
||||
|
||||
julia> d(rand(5))
|
||||
Tracked 2-element Array{Float64,1}:
|
||||
0.00257447
|
||||
-0.00449443</code></pre></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/basic.jl#L43-L62">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Conv" href="#Flux.Conv"><code>Flux.Conv</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">Conv(size, in=>out)
|
||||
Conv(size, in=>out, relu)</code></pre><p>Standard convolutional layer. <code>size</code> should be a tuple like <code>(2, 2)</code>. <code>in</code> and <code>out</code> specify the number of input and output channels respectively.</p><p>Data should be stored in WHCN order. In other words, a 100×100 RGB image would be a <code>100×100×3</code> array, and a batch of 50 would be a <code>100×100×3×50</code> array.</p><p>Takes the keyword arguments <code>pad</code>, <code>stride</code> and <code>dilation</code>.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/conv.jl#L8-L19">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.MaxPool" href="#Flux.MaxPool"><code>Flux.MaxPool</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">MaxPool(k)</code></pre><p>Max pooling layer. <code>k</code> stands for the size of the window for each dimension of the input.</p><p>Takes the keyword arguments <code>pad</code> and <code>stride</code>.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/conv.jl#L55-L61">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.MeanPool" href="#Flux.MeanPool"><code>Flux.MeanPool</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">MeanPool(k)</code></pre><p>Mean pooling layer. <code>k</code> stands for the size of the window for each dimension of the input.</p><p>Takes the keyword arguments <code>pad</code> and <code>stride</code>.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/conv.jl#L77-L83">source</a></section><h2><a class="nav-anchor" id="Recurrent-Layers-1" href="#Recurrent-Layers-1">Recurrent Layers</a></h2><p>Much like the core layers above, but can be used to process sequence data (as well as other kinds of structured data).</p><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.RNN" href="#Flux.RNN"><code>Flux.RNN</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">RNN(in::Integer, out::Integer, σ = tanh)</code></pre><p>The most basic recurrent layer; essentially acts as a <code>Dense</code> layer, but with the output fed back into the input each time step.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/recurrent.jl#L105-L110">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.LSTM" href="#Flux.LSTM"><code>Flux.LSTM</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">LSTM(in::Integer, out::Integer)</code></pre><p>Long Short Term Memory recurrent layer. Behaves like an RNN but generally exhibits a longer memory span over sequences.</p><p>See <a href="http://colah.github.io/posts/2015-08-Understanding-LSTMs/">this article</a> for a good overview of the internals.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/recurrent.jl#L150-L158">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.GRU" href="#Flux.GRU"><code>Flux.GRU</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">GRU(in::Integer, out::Integer)</code></pre><p>Gated Recurrent Unit layer. Behaves like an RNN but generally exhibits a longer memory span over sequences.</p><p>See <a href="http://colah.github.io/posts/2015-08-Understanding-LSTMs/">this article</a> for a good overview of the internals.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/recurrent.jl#L191-L199">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Recur" href="#Flux.Recur"><code>Flux.Recur</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">Recur(cell)</code></pre><p><code>Recur</code> takes a recurrent cell and makes it stateful, managing the hidden state in the background. <code>cell</code> should be a model of the form:</p><pre><code class="language-none">h, y = cell(h, x...)</code></pre><p>For example, here's a recurrent network that keeps a running total of its inputs.</p><pre><code class="language-julia">accum(h, x) = (h+x, x)
|
||||
-0.00449443</code></pre></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/basic.jl#L43-L62">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Conv" href="#Flux.Conv"><code>Flux.Conv</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">Conv(size, in=>out)
|
||||
Conv(size, in=>out, relu)</code></pre><p>Standard convolutional layer. <code>size</code> should be a tuple like <code>(2, 2)</code>. <code>in</code> and <code>out</code> specify the number of input and output channels respectively.</p><p>Data should be stored in WHCN order. In other words, a 100×100 RGB image would be a <code>100×100×3</code> array, and a batch of 50 would be a <code>100×100×3×50</code> array.</p><p>Takes the keyword arguments <code>pad</code>, <code>stride</code> and <code>dilation</code>.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/conv.jl#L8-L19">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.MaxPool" href="#Flux.MaxPool"><code>Flux.MaxPool</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">MaxPool(k)</code></pre><p>Max pooling layer. <code>k</code> stands for the size of the window for each dimension of the input.</p><p>Takes the keyword arguments <code>pad</code> and <code>stride</code>.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/conv.jl#L55-L61">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.MeanPool" href="#Flux.MeanPool"><code>Flux.MeanPool</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">MeanPool(k)</code></pre><p>Mean pooling layer. <code>k</code> stands for the size of the window for each dimension of the input.</p><p>Takes the keyword arguments <code>pad</code> and <code>stride</code>.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/conv.jl#L77-L83">source</a></section><h2><a class="nav-anchor" id="Recurrent-Layers-1" href="#Recurrent-Layers-1">Recurrent Layers</a></h2><p>Much like the core layers above, but can be used to process sequence data (as well as other kinds of structured data).</p><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.RNN" href="#Flux.RNN"><code>Flux.RNN</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">RNN(in::Integer, out::Integer, σ = tanh)</code></pre><p>The most basic recurrent layer; essentially acts as a <code>Dense</code> layer, but with the output fed back into the input each time step.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/recurrent.jl#L105-L110">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.LSTM" href="#Flux.LSTM"><code>Flux.LSTM</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">LSTM(in::Integer, out::Integer)</code></pre><p>Long Short Term Memory recurrent layer. Behaves like an RNN but generally exhibits a longer memory span over sequences.</p><p>See <a href="http://colah.github.io/posts/2015-08-Understanding-LSTMs/">this article</a> for a good overview of the internals.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/recurrent.jl#L150-L158">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.GRU" href="#Flux.GRU"><code>Flux.GRU</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">GRU(in::Integer, out::Integer)</code></pre><p>Gated Recurrent Unit layer. Behaves like an RNN but generally exhibits a longer memory span over sequences.</p><p>See <a href="http://colah.github.io/posts/2015-08-Understanding-LSTMs/">this article</a> for a good overview of the internals.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/recurrent.jl#L191-L199">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Recur" href="#Flux.Recur"><code>Flux.Recur</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">Recur(cell)</code></pre><p><code>Recur</code> takes a recurrent cell and makes it stateful, managing the hidden state in the background. <code>cell</code> should be a model of the form:</p><pre><code class="language-none">h, y = cell(h, x...)</code></pre><p>For example, here's a recurrent network that keeps a running total of its inputs.</p><pre><code class="language-julia">accum(h, x) = (h+x, x)
|
||||
rnn = Flux.Recur(accum, 0)
|
||||
rnn(2) # 2
|
||||
rnn(3) # 3
|
||||
rnn.state # 5
|
||||
rnn.(1:10) # apply to a sequence
|
||||
rnn.state # 60</code></pre></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/recurrent.jl#L7-L26">source</a></section><h2><a class="nav-anchor" id="Activation-Functions-1" href="#Activation-Functions-1">Activation Functions</a></h2><p>Non-linearities that go between layers of your model. Most of these functions are defined in <a href="https://github.com/FluxML/NNlib.jl">NNlib</a> but are available by default in Flux.</p><p>Note that, unless otherwise stated, activation functions operate on scalars. To apply them to an array you can call <code>σ.(xs)</code>, <code>relu.(xs)</code> and so on.</p><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="NNlib.σ" href="#NNlib.σ"><code>NNlib.σ</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">σ(x) = 1 / (1 + exp(-x))</code></pre><p>Classic <a href="https://en.wikipedia.org/wiki/Sigmoid_function">sigmoid</a> activation function.</p></div></div></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="NNlib.relu" href="#NNlib.relu"><code>NNlib.relu</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">relu(x) = max(0, x)</code></pre><p><a href="https://en.wikipedia.org/wiki/Rectifier_(neural_networks)">Rectified Linear Unit</a> activation function.</p></div></div></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="NNlib.leakyrelu" href="#NNlib.leakyrelu"><code>NNlib.leakyrelu</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">leakyrelu(x) = max(0.01x, x)</code></pre><p>Leaky <a href="https://en.wikipedia.org/wiki/Rectifier_(neural_networks)">Rectified Linear Unit</a> activation function. You can also specify the coefficient explicitly, e.g. <code>leakyrelu(x, 0.01)</code>.</p></div></div></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="NNlib.elu" href="#NNlib.elu"><code>NNlib.elu</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">elu(x, α = 1) =
|
||||
rnn.state # 60</code></pre></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/recurrent.jl#L7-L26">source</a></section><h2><a class="nav-anchor" id="Activation-Functions-1" href="#Activation-Functions-1">Activation Functions</a></h2><p>Non-linearities that go between layers of your model. Most of these functions are defined in <a href="https://github.com/FluxML/NNlib.jl">NNlib</a> but are available by default in Flux.</p><p>Note that, unless otherwise stated, activation functions operate on scalars. To apply them to an array you can call <code>σ.(xs)</code>, <code>relu.(xs)</code> and so on.</p><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="NNlib.σ" href="#NNlib.σ"><code>NNlib.σ</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">σ(x) = 1 / (1 + exp(-x))</code></pre><p>Classic <a href="https://en.wikipedia.org/wiki/Sigmoid_function">sigmoid</a> activation function.</p></div></div></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="NNlib.relu" href="#NNlib.relu"><code>NNlib.relu</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">relu(x) = max(0, x)</code></pre><p><a href="https://en.wikipedia.org/wiki/Rectifier_(neural_networks)">Rectified Linear Unit</a> activation function.</p></div></div></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="NNlib.leakyrelu" href="#NNlib.leakyrelu"><code>NNlib.leakyrelu</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">leakyrelu(x) = max(0.01x, x)</code></pre><p>Leaky <a href="https://en.wikipedia.org/wiki/Rectifier_(neural_networks)">Rectified Linear Unit</a> activation function. You can also specify the coefficient explicitly, e.g. <code>leakyrelu(x, 0.01)</code>.</p></div></div></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="NNlib.elu" href="#NNlib.elu"><code>NNlib.elu</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">elu(x, α = 1) =
|
||||
x > 0 ? x : α * (exp(x) - 1)</code></pre><p>Exponential Linear Unit activation function. See <a href="https://arxiv.org/abs/1511.07289">Fast and Accurate Deep Network Learning by Exponential Linear Units</a>. You can also specify the coefficient explicitly, e.g. <code>elu(x, 1)</code>.</p></div></div></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="NNlib.swish" href="#NNlib.swish"><code>NNlib.swish</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">swish(x) = x * σ(x)</code></pre><p>Self-gated actvation function. See <a href="https://arxiv.org/pdf/1710.05941.pdf">Swish: a Self-Gated Activation Function</a>.</p></div></div></section><h2><a class="nav-anchor" id="Normalisation-and-Regularisation-1" href="#Normalisation-and-Regularisation-1">Normalisation & Regularisation</a></h2><p>These layers don't affect the structure of the network but may improve training times or reduce overfitting.</p><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.testmode!" href="#Flux.testmode!"><code>Flux.testmode!</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">testmode!(m)
|
||||
testmode!(m, false)</code></pre><p>Put layers like <a href="layers.html#Flux.Dropout"><code>Dropout</code></a> and <a href="layers.html#Flux.BatchNorm"><code>BatchNorm</code></a> into testing mode (or back to training mode with <code>false</code>).</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/normalise.jl#L1-L7">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.BatchNorm" href="#Flux.BatchNorm"><code>Flux.BatchNorm</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">BatchNorm(channels::Integer, σ = identity;
|
||||
testmode!(m, false)</code></pre><p>Put layers like <a href="layers.html#Flux.Dropout"><code>Dropout</code></a> and <a href="layers.html#Flux.BatchNorm"><code>BatchNorm</code></a> into testing mode (or back to training mode with <code>false</code>).</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/normalise.jl#L1-L7">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.BatchNorm" href="#Flux.BatchNorm"><code>Flux.BatchNorm</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">BatchNorm(channels::Integer, σ = identity;
|
||||
initβ = zeros, initγ = ones,
|
||||
ϵ = 1e-8, momentum = .1)</code></pre><p>Batch Normalization layer. The <code>channels</code> input should be the size of the channel dimension in your data (see below).</p><p>Given an array with <code>N</code> dimensions, call the <code>N-1</code>th the channel dimension. (For a batch of feature vectors this is just the data dimension, for <code>WHCN</code> images it's the usual channel dimension.)</p><p><code>BatchNorm</code> computes the mean and variance for each each <code>W×H×1×N</code> slice and shifts them to have a new mean and variance (corresponding to the learnable, per-channel <code>bias</code> and <code>scale</code> parameters).</p><p>See <a href="https://arxiv.org/pdf/1502.03167.pdf">Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift</a>.</p><p>Example:</p><pre><code class="language-julia">m = Chain(
|
||||
Dense(28^2, 64),
|
||||
BatchNorm(64, relu),
|
||||
Dense(64, 10),
|
||||
BatchNorm(10),
|
||||
softmax)</code></pre></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/normalise.jl#L69-L98">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Dropout" href="#Flux.Dropout"><code>Flux.Dropout</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">Dropout(p)</code></pre><p>A Dropout layer. For each input, either sets that input to <code>0</code> (with probability <code>p</code>) or scales it by <code>1/(1-p)</code>. This is used as a regularisation, i.e. it reduces overfitting during training.</p><p>Does nothing to the input once in <a href="layers.html#Flux.testmode!"><code>testmode!</code></a>.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/normalise.jl#L15-L23">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.LayerNorm" href="#Flux.LayerNorm"><code>Flux.LayerNorm</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">LayerNorm(h::Integer)</code></pre><p>A <a href="https://arxiv.org/pdf/1607.06450.pdf">normalisation layer</a> designed to be used with recurrent hidden states of size <code>h</code>. Normalises the mean/stddev of each input before applying a per-neuron gain/bias.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/layers/normalise.jl#L46-L53">source</a></section><footer><hr/><a class="previous" href="regularisation.html"><span class="direction">Previous</span><span class="title">Regularisation</span></a><a class="next" href="../training/optimisers.html"><span class="direction">Next</span><span class="title">Optimisers</span></a></footer></article></body></html>
|
||||
softmax)</code></pre></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/normalise.jl#L69-L98">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Dropout" href="#Flux.Dropout"><code>Flux.Dropout</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">Dropout(p)</code></pre><p>A Dropout layer. For each input, either sets that input to <code>0</code> (with probability <code>p</code>) or scales it by <code>1/(1-p)</code>. This is used as a regularisation, i.e. it reduces overfitting during training.</p><p>Does nothing to the input once in <a href="layers.html#Flux.testmode!"><code>testmode!</code></a>.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/normalise.jl#L15-L23">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.LayerNorm" href="#Flux.LayerNorm"><code>Flux.LayerNorm</code></a> — <span class="docstring-category">Type</span>.</div><div><div><pre><code class="language-none">LayerNorm(h::Integer)</code></pre><p>A <a href="https://arxiv.org/pdf/1607.06450.pdf">normalisation layer</a> designed to be used with recurrent hidden states of size <code>h</code>. Normalises the mean/stddev of each input before applying a per-neuron gain/bias.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/layers/normalise.jl#L46-L53">source</a></section><footer><hr/><a class="previous" href="regularisation.html"><span class="direction">Previous</span><span class="title">Regularisation</span></a><a class="next" href="../training/optimisers.html"><span class="direction">Next</span><span class="title">Optimisers</span></a></footer></article></body></html>
|
||||
|
@ -541,7 +541,7 @@ var documenterSearchIndex = {"docs": [
|
||||
"page": "Backpropagation",
|
||||
"title": "Custom Gradients",
|
||||
"category": "section",
|
||||
"text": "We can hook in to the processes above to implement custom gradients for a function or kernel. For a toy example, imagine a custom implementation of minus:minus(a, b) = a - bFirstly, we must tell the tracker system to stop when it sees a call to minus, and record it. We can do this using dispatch:using Flux.Tracker: TrackedReal, track, @grad\n\nminus(a::TrackedArray, b::TrackedArray) = Tracker.track(minus, a, b)track takes care of building a new Tracked object and recording the operation on the tape. We just need to provide a gradient definition.@grad function minus(a, b)\n return minus(data(a),data(b)), Δ -> (Δ, -Δ)\nendThis is essentially just a way of overloading the forward function we saw above. We strip tracking from a and b so that we are calling the original definition of minus (otherwise, we\'d just try to track the call again and hit an infinite regress).Note that in the backpropagator we don\'t call data(a); we do in fact want to track this, since nest AD will take a derivative through the backpropagator itself. For example, the gradient of * might look like this.@grad a * b = data(a)*data(b), Δ -> (Δ*b, a*Δ)For multi-argument functions with custom gradients, you likely want to catch not just minus(::TrackedArray, ::TrackedArray) but also minus(::Array, TrackedArray) and so on. To do so, just define those extra signatures as needed:minus(a::AbstractArray, b::TrackedArray) = Tracker.track(minus, a, b)\nminus(a::TrackedArray, b::AbstractArray) = Tracker.track(minus, a, b)"
|
||||
"text": "We can hook in to the processes above to implement custom gradients for a function or kernel. For a toy example, imagine a custom implementation of minus:minus(a, b) = a - bFirstly, we must tell the tracker system to stop when it sees a call to minus, and record it. We can do this using dispatch:using Flux.Tracker: TrackedArray, track, @grad\n\nminus(a::TrackedArray, b::TrackedArray) = track(minus, a, b)track takes care of building a new Tracked object and recording the operation on the tape. We just need to provide a gradient definition.@grad function minus(a, b)\n return minus(data(a), data(b)), Δ -> (Δ, -Δ)\nendThis is essentially just a way of overloading the forward function we saw above. We strip tracking from a and b so that we are calling the original definition of minus (otherwise, we\'d just try to track the call again and hit an infinite regress).Note that in the backpropagator we don\'t call data(a); we do in fact want to track this, since nest AD will take a derivative through the backpropagator itself. For example, the gradient of * might look like this.@grad a * b = data(a)*data(b), Δ -> (Δ*b, a*Δ)We can then calculate the first derivative of minus as follows:a = param([1,2,3])\nb = param([3,2,1])\n\nc = minus(a, b) # [-2.0 (tracked), 0.0 (tracked), 2.0 (tracked)]\n\nTracker.back!(c, 1)\nTracker.grad(a) # [1.00, 1.00, 1.00]\nTracker.grad(b) # [-1.00, -1.00, -1.00]For multi-argument functions with custom gradients, you likely want to catch not just minus(::TrackedArray, ::TrackedArray) but also minus(::Array, TrackedArray) and so on. To do so, just define those extra signatures as needed:minus(a::AbstractArray, b::TrackedArray) = Tracker.track(minus, a, b)\nminus(a::TrackedArray, b::AbstractArray) = Tracker.track(minus, a, b)"
|
||||
},
|
||||
|
||||
{
|
||||
|
@ -29,4 +29,4 @@ end</code></pre><p>If we call <code>sgd</code>, the parameters <code>W</code> an
|
||||
Dense(10, 5, σ),
|
||||
Dense(5, 2), softmax)</code></pre><p>Instead of having to write <code>[m[1].W, m[1].b, ...]</code>, Flux provides a params function <code>params(m)</code> that returns a list of all parameters in the model for you.</p><p>For the update step, there's nothing whatsoever wrong with writing the loop above – it'll work just fine – but Flux provides various <em>optimisers</em> that make it more convenient.</p><pre><code class="language-julia">opt = SGD([W, b], 0.1) # Gradient descent with learning rate 0.1
|
||||
|
||||
opt() # Carry out the update, modifying `W` and `b`.</code></pre><p>An optimiser takes a parameter list and returns a function that does the same thing as <code>update</code> above. We can pass either <code>opt</code> or <code>update</code> to our <a href="training.html">training loop</a>, which will then run the optimiser after every mini-batch of data.</p><h2><a class="nav-anchor" id="Optimiser-Reference-1" href="#Optimiser-Reference-1">Optimiser Reference</a></h2><p>All optimisers return a function that, when called, will update the parameters passed to it.</p><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Optimise.SGD" href="#Flux.Optimise.SGD"><code>Flux.Optimise.SGD</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">SGD(params, η = 0.1; decay = 0)</code></pre><p>Classic gradient descent optimiser with learning rate <code>η</code>. For each parameter <code>p</code> and its gradient <code>δp</code>, this runs <code>p -= η*δp</code>.</p><p>Supports inverse decaying learning rate if the <code>decay</code> argument is provided.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/optimise/interface.jl#L14-L21">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Optimise.Momentum" href="#Flux.Optimise.Momentum"><code>Flux.Optimise.Momentum</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">Momentum(params, η = 0.01; ρ = 0.9, decay = 0)</code></pre><p>SGD with learning rate <code>η</code>, momentum <code>ρ</code> and optional learning rate inverse decay.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/optimise/interface.jl#L25-L29">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Optimise.Nesterov" href="#Flux.Optimise.Nesterov"><code>Flux.Optimise.Nesterov</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">Nesterov(params, η = 0.01; ρ = 0.9, decay = 0)</code></pre><p>SGD with learning rate <code>η</code>, Nesterov momentum <code>ρ</code> and optional learning rate inverse decay.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/optimise/interface.jl#L33-L37">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Optimise.ADAM" href="#Flux.Optimise.ADAM"><code>Flux.Optimise.ADAM</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">ADAM(params, η = 0.001; β1 = 0.9, β2 = 0.999, ϵ = 1e-08, decay = 0)</code></pre><p><a href="https://arxiv.org/abs/1412.6980v8">ADAM</a> optimiser.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/5d3cc044cd81506b69f9f228f5548fdbc96269bd/src/optimise/interface.jl#L51-L55">source</a></section><footer><hr/><a class="previous" href="../models/layers.html"><span class="direction">Previous</span><span class="title">Model Reference</span></a><a class="next" href="training.html"><span class="direction">Next</span><span class="title">Training</span></a></footer></article></body></html>
|
||||
opt() # Carry out the update, modifying `W` and `b`.</code></pre><p>An optimiser takes a parameter list and returns a function that does the same thing as <code>update</code> above. We can pass either <code>opt</code> or <code>update</code> to our <a href="training.html">training loop</a>, which will then run the optimiser after every mini-batch of data.</p><h2><a class="nav-anchor" id="Optimiser-Reference-1" href="#Optimiser-Reference-1">Optimiser Reference</a></h2><p>All optimisers return a function that, when called, will update the parameters passed to it.</p><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Optimise.SGD" href="#Flux.Optimise.SGD"><code>Flux.Optimise.SGD</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">SGD(params, η = 0.1; decay = 0)</code></pre><p>Classic gradient descent optimiser with learning rate <code>η</code>. For each parameter <code>p</code> and its gradient <code>δp</code>, this runs <code>p -= η*δp</code>.</p><p>Supports inverse decaying learning rate if the <code>decay</code> argument is provided.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/optimise/interface.jl#L14-L21">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Optimise.Momentum" href="#Flux.Optimise.Momentum"><code>Flux.Optimise.Momentum</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">Momentum(params, η = 0.01; ρ = 0.9, decay = 0)</code></pre><p>SGD with learning rate <code>η</code>, momentum <code>ρ</code> and optional learning rate inverse decay.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/optimise/interface.jl#L25-L29">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Optimise.Nesterov" href="#Flux.Optimise.Nesterov"><code>Flux.Optimise.Nesterov</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">Nesterov(params, η = 0.01; ρ = 0.9, decay = 0)</code></pre><p>SGD with learning rate <code>η</code>, Nesterov momentum <code>ρ</code> and optional learning rate inverse decay.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/optimise/interface.jl#L33-L37">source</a></section><section class="docstring"><div class="docstring-header"><a class="docstring-binding" id="Flux.Optimise.ADAM" href="#Flux.Optimise.ADAM"><code>Flux.Optimise.ADAM</code></a> — <span class="docstring-category">Function</span>.</div><div><div><pre><code class="language-none">ADAM(params, η = 0.001; β1 = 0.9, β2 = 0.999, ϵ = 1e-08, decay = 0)</code></pre><p><a href="https://arxiv.org/abs/1412.6980v8">ADAM</a> optimiser.</p></div></div><a class="source-link" target="_blank" href="https://github.com/FluxML/Flux.jl/blob/3285afa45a7f848e56fe4ebf910295f5c1069244/src/optimise/interface.jl#L51-L55">source</a></section><footer><hr/><a class="previous" href="../models/layers.html"><span class="direction">Previous</span><span class="title">Model Reference</span></a><a class="next" href="training.html"><span class="direction">Next</span><span class="title">Training</span></a></footer></article></body></html>
|
||||
|
Loading…
Reference in New Issue
Block a user