Merge branch 'master' into cat-fix

README.md

@@ -1,6 +1,6 @@
 # Флукс
 
-[![Build Status](https://travis-ci.org/FluxML/Flux.jl.svg?branch=master)](https://travis-ci.org/FluxML/Flux.jl) [![](https://img.shields.io/badge/docs-stable-blue.svg)](https://fluxml.github.io/Flux.jl/stable/) [![Join the chat at https://gitter.im/FluxML](https://badges.gitter.im/FluxML/Lobby.svg)](https://gitter.im/FluxML/Lobby) [Slack](https://discourse.julialang.org/t/announcing-a-julia-slack/4866)
+[![Build Status](https://travis-ci.org/FluxML/Flux.jl.svg?branch=master)](https://travis-ci.org/FluxML/Flux.jl) [![](https://img.shields.io/badge/docs-stable-blue.svg)](https://fluxml.github.io/Flux.jl/stable/) [![Join the chat at https://gitter.im/FluxML](https://badges.gitter.im/FluxML/Lobby.svg)](https://gitter.im/FluxML/Lobby) [Slack](https://slackinvite.julialang.org/)
 
 Flux is a refreshing approach to machine learning. It provides lightweight abstractions on top of Julia's native GPU and AD support, while remaining fully hackable (right down to the [GPU kernels](https://github.com/FluxML/CuArrays.jl)).
 

REQUIRE

@@ -3,5 +3,6 @@ DataFlow 0.2.1
 Juno
 MacroTools 0.3.3
 NNlib
-ForwardDiff
+ForwardDiff 0.5.0
 Requires
+Adapt

docs/src/models/layers.md

@@ -5,6 +5,7 @@ These core layers form the foundation of almost all neural networks.
 ```@docs
 Chain
 Dense
+Conv2D
 ```
 
 ## Recurrent Layers
@@ -37,6 +38,7 @@ These layers don't affect the structure of the network but may improve training
 
 ```@docs
 Flux.testmode!
+BatchNorm
 Dropout
 LayerNorm
 ```

src/Flux.jl

@@ -7,12 +7,14 @@ module Flux
 using Juno, Requires
 using Lazy: @forward
 
-export Chain, Dense, RNN, LSTM, Dropout, LayerNorm,
+export Chain, Dense, RNN, LSTM, GRU, Conv2D,
-  SGD, ADAM, Momentum, Nesterov,
+  Dropout, LayerNorm, BatchNorm,
+  SGD, ADAM, Momentum, Nesterov, AMSGrad,
   param, params, mapleaves
 
 using NNlib
-export σ, sigmoid, relu, leakyrelu, elu, swish, softmax
+export σ, sigmoid, relu, leakyrelu, elu, swish, softmax,
+  conv2d, maxpool2d, avgpool2d
 
 include("tracker/Tracker.jl")
 using .Tracker
@@ -26,6 +28,7 @@ include("treelike.jl")
 
 include("layers/stateless.jl")
 include("layers/basic.jl")
+include("layers/conv.jl")
 include("layers/recurrent.jl")
 include("layers/normalisation.jl")
 

src/data/cmudict.jl

@@ -23,17 +23,17 @@ end
 
 function symbols()
   load()
-  Symbol.(split(readstring(deps("CMUDict", "cmudict.symbols")),
+  Symbol.(split(readstring(deps("cmudict", "cmudict.symbols")),
                 "\n", keep = false))
 end
 
 function rawdict()
   load()
   Dict(String(xs[1]) => Symbol.(xs[2:end]) for xs in
-       filter(!isempty, split.(split(readstring(deps("CMUDict", "cmudict")), "\n"))))
+       filter(!isempty, split.(split(readstring(deps("cmudict", "cmudict")), "\n"))))
 end
 
-validword(s) = ismatch(r"^[\w-\.]+$", s)
+validword(s) = ismatch(r"^[\w\-\.]+$", s)
 
 cmudict() = filter((s, ps) -> validword(s), rawdict())
 

src/layers/basic.jl

@@ -63,8 +63,10 @@ struct Dense{F,S,T}
   b::T
 end
 
-Dense(in::Integer, out::Integer, σ = identity; init = initn) =
+function Dense(in::Integer, out::Integer, σ = identity;
-  Dense(σ, param(init(out, in)), param(init(out)))
+               initW = glorot_uniform, initb = zeros)
+  return Dense(σ, param(initW(out, in)), param(initb(out)))
+end
 
 treelike(Dense)
 

src/layers/conv.jl

@@ -0,0 +1,33 @@
+"""
+    Conv2D(size, in=>out)
+    Conv2d(size, in=>out, relu)
+
+Standard convolutional layer. `size` should be a tuple like `(2, 2)`.
+`in` and `out` specify the number of input and output channels respectively.
+
+Data should be stored in HWCN order. In other words, a 100×100 RGB image would
+be a `100×100×3` array, and a batch of 50 would be a `100×100×3×50` array.
+
+Takes the keyword arguments `pad` and `stride`.
+"""
+struct Conv2D{F,A}
+  σ::F
+  weight::A
+  stride::Int
+  pad::Int
+end
+
+Conv2D(k::NTuple{2,Integer}, ch::Pair{<:Integer,<:Integer}, σ = identity;
+       init = initn, stride = 1, pad = 0) =
+  Conv2D(σ, param(init(k..., ch...)), stride, pad)
+
+Flux.treelike(Conv2D)
+
+(c::Conv2D)(x) = c.σ.(conv2d(x, c.weight, stride = c.stride, padding = c.pad))
+
+function Base.show(io::IO, l::Conv2D)
+  print(io, "Conv2D((", size(l.weight, 1), ", ", size(l.weight, 2), ")")
+  print(io, ", ", size(l.weight, 3), "=>", size(l.weight, 4))
+  l.σ == identity || print(io, ", ", l.σ)
+  print(io, ")")
+end

src/layers/normalisation.jl

@@ -2,8 +2,8 @@
     testmode!(m)
     testmode!(m, false)
 
-Put layers like [`Dropout`](@ref) and `BatchNorm` into testing mode (or back to
+Put layers like [`Dropout`](@ref) and [`BatchNorm`](@ref) into testing mode
-training mode with `false`).
+(or back to training mode with `false`).
 """
 function testmode!(m, val::Bool=true)
   prefor(x -> _testmode!(x, val), m)
@@ -45,6 +45,7 @@ end
 _testmode!(a::Dropout, test) = (a.active = !test)
 
 """
+
     LayerNorm(h::Integer)
 
 A [normalisation layer](https://arxiv.org/pdf/1607.06450.pdf) designed to be
@@ -65,3 +66,77 @@ treelike(LayerNorm)
 function Base.show(io::IO, l::LayerNorm)
   print(io, "LayerNorm(", length(l.diag.α), ")")
 end
+
+"""
+    BatchNorm(dims...; λ = identity,
+              initβ = zeros, initγ = ones, ϵ = 1e-8, momentum = .1)
+
+Batch Normalization Layer for [`Dense`](@ref) layer.
+
+See [Batch Normalization: Accelerating Deep Network Training by Reducing
+     Internal Covariate Shift](https://arxiv.org/pdf/1502.03167.pdf)
+
+In the example of MNIST,
+in order to normalize the input of other layer,
+put the `BatchNorm` layer before activation function.
+
+```julia
+m = Chain(
+  Dense(28^2, 64),
+  BatchNorm(64, λ = relu),
+  Dense(64, 10),
+  BatchNorm(10),
+  softmax)
+```
+"""
+mutable struct BatchNorm{F,V,N}
+  λ::F  # activation function
+  β::V  # bias
+  γ::V  # scale
+  μ     # moving mean
+  σ     # moving std
+  ϵ::N
+  momentum::N
+  active::Bool
+end
+
+BatchNorm(dims::Integer...; λ = identity,
+          initβ = zeros, initγ = ones, ϵ = 1e-8, momentum = .1) =
+  BatchNorm(λ, param(initβ(dims)), param(initγ(dims)), 0., 1., ϵ, momentum, true)
+
+function (BN::BatchNorm)(x)
+  λ, γ, β = BN.λ, BN.γ, BN.β
+
+  if !BN.active
+    μ = BN.μ
+    σ = BN.σ
+  else
+    T = eltype(x)
+
+    ϵ = T(BN.ϵ)
+    m = size(x, 2)  # batch size
+    μ = mean(x, 2)
+    σ = sqrt.(sum((x .- μ).^2, 2) ./ m .+ ϵ)
+
+    # update moving mean/std
+    mtm = T(BN.momentum)
+    BN.μ = (1 - mtm) .* BN.μ .+ mtm .* μ.data
+    BN.σ = (1 - mtm) .* BN.σ .+ mtm .* σ.data .* m ./ (m - 1)
+  end
+
+  λ.(γ .* ((x .- μ) ./ σ) .+ β)
+end
+
+children(BN::BatchNorm) =
+  (BN.λ, BN.β, BN.γ, BN.μ, BN.σ, BN.momentum, BN.ϵ, BN.active)
+
+mapchildren(f, BN::BatchNorm) =  # e.g. mapchildren(cu, BN)
+  BatchNorm(BN.λ, f(BN.β), f(BN.γ), BN.μ, BN.σ, BN.momentum, BN.ϵ, BN.active)
+
+_testmode!(BN::BatchNorm, test) = (BN.active = !test)
+
+function Base.show(io::IO, l::BatchNorm)
+  print(io, "BatchNorm($(join(size(l.β), ", "))")
+  (l.λ == identity) || print(io, ", λ = $(l.λ)")
+  print(io, ")")
+end

src/layers/recurrent.jl

@@ -79,8 +79,8 @@ struct RNNCell{D,V}
   h::V
 end
 
-RNNCell(in::Integer, out::Integer, σ = tanh; init = initn) =
+RNNCell(in::Integer, out::Integer, σ = tanh; initW = glorot_uniform, initb = zeros) =
-  RNNCell(Dense(in+out, out, σ, init = init), param(init(out)))
+  RNNCell(Dense(in+out, out, σ, initW = initW, initb = initb), param(initW(out)))
 
 function (m::RNNCell)(h, x)
   h = m.d(combine(x, h))
@@ -113,10 +113,10 @@ struct LSTMCell{D1,D2,V}
   h::V; c::V
 end
 
-function LSTMCell(in, out; init = initn)
+function LSTMCell(in, out; initW = glorot_uniform, initb = zeros)
-  cell = LSTMCell([Dense(in+out, out, σ, init = init) for _ = 1:3]...,
+  cell = LSTMCell([Dense(in+out, out, σ, initW = initW, initb = initb) for _ = 1:3]...,
-                  Dense(in+out, out, tanh, init = init),
+                  Dense(in+out, out, tanh, initW = initW, initb = initb),
-                  param(init(out)), param(init(out)))
+                  param(initW(out)), param(initW(out)))
   cell.forget.b.data .= 1
   return cell
 end
@@ -150,3 +150,49 @@ See [this article](http://colah.github.io/posts/2015-08-Understanding-LSTMs/)
 for a good overview of the internals.
 """
 LSTM(a...; ka...) = Recur(LSTMCell(a...; ka...))
+
+# GRU
+
+struct GRUCell{D1,D2,V}
+  update::D1
+  reset::D1
+  candidate::D2
+  h::V
+end
+
+function GRUCell(in, out)
+  cell = GRUCell(Dense(in+out, out, σ),
+                 Dense(in+out, out, σ),
+                 Dense(in+out, out, tanh),
+                 param(initn(out)))
+  return cell
+end
+
+function (m::GRUCell)(h, x)
+  x′   = combine(x, h)
+  z    = m.update(x′)
+  r    = m.reset(x′)
+  h̃    = m.candidate(combine(r.*h, x))
+  h = (1.-z).*h .+ z.*h̃
+  return h, h
+end
+
+hidden(m::GRUCell) = m.h
+
+treelike(GRUCell)
+
+Base.show(io::IO, m::GRUCell) =
+  print(io, "GRUCell(",
+        size(m.update.W, 2) - size(m.update.W, 1), ", ",
+        size(m.update.W, 1), ')')
+
+"""
+    GRU(in::Integer, out::Integer, σ = tanh)
+
+Gated Recurrent Unit layer. Behaves like an RNN but generally
+exhibits a longer memory span over sequences.
+
+See [this article](http://colah.github.io/posts/2015-08-Understanding-LSTMs/)
+for a good overview of the internals.
+"""
+GRU(a...; ka...) = Recur(GRUCell(a...; ka...))

src/layers/stateless.jl

@@ -4,8 +4,9 @@ using NNlib: log_fast
 
 mse(ŷ, y) = sum((ŷ .- y).^2)/length(y)
 
-crossentropy(ŷ::AbstractVecOrMat, y::AbstractVecOrMat) =
+function crossentropy(ŷ::AbstractVecOrMat, y::AbstractVecOrMat; weight = 1)
-  -sum(y .* log_fast.(ŷ)) / size(y, 2)
+  return -sum(y .* log_fast.(ŷ) .* weight) / size(y, 2)
+end
 
 @deprecate logloss(x, y) crossentropy(x, y)
 

src/onehot.jl

@@ -18,7 +18,9 @@ end
 
 Base.size(xs::OneHotMatrix) = (Int64(xs.height),length(xs.data))
 
-Base.getindex(xs::OneHotMatrix, i::Int, j::Int) = xs.data[j][i]
+Base.getindex(xs::OneHotMatrix, i::Integer, j::Integer) = xs.data[j][i]
+Base.getindex(xs::OneHotMatrix, ::Colon, i::Integer) = xs.data[i]
+Base.getindex(xs::OneHotMatrix, ::Colon, i::AbstractArray) = OneHotMatrix(xs.height, xs.data[i])
 
 A::AbstractMatrix * B::OneHotMatrix = A[:, map(x->x.ix, B.data)]
 
@@ -26,7 +28,7 @@ Base.hcat(x::OneHotVector, xs::OneHotVector...) = OneHotMatrix(length(x), [x, xs
 
 batch(xs::AbstractArray{<:OneHotVector}) = OneHotMatrix(length(first(xs)), xs)
 
-import NNlib.adapt
+import Adapt.adapt
 
 adapt(T, xs::OneHotMatrix) = OneHotMatrix(xs.height, adapt(T, xs.data))
 

src/optimise/Optimise.jl

@@ -1,7 +1,7 @@
 module Optimise
 
 export update!, params, train!,
-  SGD, ADAM, Momentum, Nesterov, RMSProp, ADAGrad, ADADelta
+  SGD, ADAM, Momentum, Nesterov, RMSProp, ADAGrad, ADADelta, AMSGrad
 
 struct Param{T}
   x::T

src/optimise/interface.jl

@@ -1,5 +1,7 @@
 call(f, xs...) = f(xs...)
 
+# note for optimisers: set to zero
+# p.Δ at the end of the weigths update
 function optimiser(ps, fs...)
   ps = [Param(p) for p in ps]
   fs = map(ps) do p
@@ -10,64 +12,73 @@ function optimiser(ps, fs...)
 end
 
 """
-    SGD(params, η = 1; decay = 0)
+    SGD(params, η = 0.1; decay = 0)
 
-Classic gradient descent optimiser. For each parameter `p` and its
+Classic gradient descent optimiser with learning rate `η`.
-gradient `δp`, this runs `p -= η*δp`.
+For each parameter `p` and its gradient `δp`, this runs `p -= η*δp`.
 
-Supports decayed learning rate decay if the `decay` argument is provided.
+Supports inverse decaying learning rate if the `decay` argument is provided.
 """
-SGD(ps, η = 1; decay = 0) =
+SGD(ps, η = 0.1; decay = 0) =
-  optimiser(ps, p -> invdecay(p, decay), p -> descent(p, η))
+  optimiser(ps, p -> invdecay(p, decay), p -> descent(p,η))
 
 """
-    Momentum(params, ρ, decay = 0)
+    Momentum(params, η = 0.01; ρ = 0.9, decay = 0)
 
-SGD with momentum `ρ` and optional learning rate decay.
+SGD with learning rate  `η`, momentum `ρ` and optional learning rate inverse decay.
 """
-Momentum(ps, ρ; decay = 0) =
+Momentum(ps, η = 0.01; ρ = 0.9, decay = 0) =
-  optimiser(ps, p -> momentum(p, ρ), p -> invdecay(p, decay), p -> descent(p, 1))
+  optimiser(ps, p->invdecay(p,decay), p->momentum(p, ρ, η), p->descent(p,1))
 
 """
-    Nesterov(params, ρ, decay = 0)
+    Nesterov(params, η = 0.01; ρ = 0.9, decay = 0)
 
-SGD with Nesterov momentum `ρ` and optional learning rate decay.
+SGD with learning rate  `η`, Nesterov momentum `ρ` and optional learning rate inverse decay.
 """
-Nesterov(ps, ρ; decay = 0) =
+Nesterov(ps, η = 0.01; ρ = 0.9, decay = 0) =
-  optimiser(ps, p -> nesterov(p, ρ), p -> invdecay(p, decay), p -> descent(p, 1))
+  optimiser(ps, p->invdecay(p,decay), p->nesterov(p, ρ, η), p->descent(p,1))
 
 """
-    RMSProp(params; η = 0.001, ρ = 0.9, ϵ = 1e-8, decay = 0)
+    RMSProp(params, η = 0.001; ρ = 0.9, ϵ = 1e-8, decay = 0)
 
 [RMSProp](http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf)
 optimiser. Parameters other than learning rate don't need tuning. Often a good
 choice for recurrent networks.
 """
 RMSProp(ps, η = 0.001; ρ = 0.9, ϵ = 1e-8, decay = 0) =
-  optimiser(ps, p -> rmsprop(p; η = η, ρ = ρ, ϵ = ϵ), p -> invdecay(p, decay), p -> descent(p, 1))
+  optimiser(ps, p->rmsprop(p; η=η, ρ=ρ, ϵ=ϵ), p->invdecay(p,decay), p->descent(p,1))
 
 """
-    ADAM(params; η = 0.001, β1 = 0.9, β2 = 0.999, ϵ = 1e-08, decay = 0)
+    ADAM(params, η = 0.001; β1 = 0.9, β2 = 0.999, ϵ = 1e-08, decay = 0)
 
 [ADAM](https://arxiv.org/abs/1412.6980v8) optimiser.
 """
 ADAM(ps, η = 0.001; β1 = 0.9, β2 = 0.999, ϵ = 1e-08, decay = 0) =
-  optimiser(ps, p -> adam(p; η = η, β1 = β1, β2 = β2, ϵ = ϵ), p -> invdecay(p, decay), p -> descent(p, 1))
+  optimiser(ps, p->adam(p; η=η, β1=β1, β2=β2, ϵ=ϵ), p->invdecay(p,decay), p->descent(p,1))
 
 """
-    ADAGrad(params; η = 0.01, ϵ = 1e-8, decay = 0)
+    ADAGrad(params, η = 0.01; ϵ = 1e-8, decay = 0)
 
 [ADAGrad](http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) optimiser.
 Parameters don't need tuning.
 """
-ADAGrad(ps; η = 0.01, ϵ = 1e-8, decay = 0) =
+ADAGrad(ps, η = 0.01; ϵ = 1e-8, decay = 0) =
-  optimiser(ps, p -> adagrad(p; η = η, ϵ = ϵ), p -> invdecay(p, decay), p -> descent(p, 1))
+  optimiser(ps, p->adagrad(p; η=η, ϵ=ϵ), p->invdecay(p,decay), p->descent(p,1))
 
 """
-    ADADelta(params; η = 0.01, ρ = 0.95, ϵ = 1e-8, decay = 0)
+    ADADelta(params; ρ = 0.9, ϵ = 1e-8, decay = 0)
 
 [ADADelta](http://arxiv.org/abs/1212.5701) optimiser. Parameters don't need
 tuning.
 """
-ADADelta(ps; η = 0.01, ρ = 0.95, ϵ = 1e-8, decay = 0) =
+ADADelta(ps; ρ = 0.9, ϵ = 1e-8, decay = 0) =
-  optimiser(ps, p -> adadelta(p; ρ = ρ, ϵ = ϵ), p -> invdecay(p, decay), p -> descent(p, 1))
+  optimiser(ps, p->adadelta(p; ρ=ρ, ϵ=ϵ), p->descent(p,1))
+
+"""
+    AMSGrad(params; η = 0.001, β1 = 0.9, β2 = 0.999, ϵ = 1e-08, decay = 0)
+
+[AMSGrad](https://openreview.net/forum?id=ryQu7f-RZ) optimiser. Parameters don't need
+tuning.
+"""
+AMSGrad(ps, η = 0.001; β1 = 0.9, β2 = 0.999, ϵ = 1e-08, decay = 0) =
+  optimiser(ps, p -> amsgrad(p; η = η, β1 = β1, β2 = β2, ϵ = ϵ), p -> invdecay(p, decay), p -> descent(p, 1))

src/optimise/optimisers.jl

@@ -1,74 +1,97 @@
 function descent(p::Param, η::Real)
   function ()
-    p.x .-= p.Δ .* η
+    @. p.x -= η * p.Δ
-    p.Δ .= 0
+    @. p.Δ = 0
   end
 end
 
-function momentum(p::Param, ρ::Real)
+function momentum(p::Param, ρ, η)
-  mo = zeros(p.x)
+  v = zeros(p.x)
-  () -> p.Δ .= mo .= ρ .* mo .+ p.Δ
-end
-
-function nesterov(p::Param, ρ::Real)
-  mo = zeros(p.x)
   function ()
-    mo  .= ρ .* mo .+ p.Δ
+    @. v = ρ * v - η * p.Δ
-    p.Δ .= ρ .* mo .+ p.Δ
+    @. p.Δ = -v
   end
 end
 
-function clip(p::Param, thresh::Real)
+# Ref. https://arxiv.org/pdf/1212.0901.pdf
-  () -> clamp!(p.Δ, -thresh, thresh)
+function nesterov(p::Param, ρ, η)
-end
+  v = zeros(p.x)
-
-function weightdecay(p::Param, γ::Real)
-  () -> p.Δ .+= γ .* p.x
-end
-
-function invdecay(p::Param, γ::Real)
-  n = 0
   function ()
-    p.Δ .*= 1 / (1 + γ * n)
+    d = @. ρ^2 * v - (1+ρ) * η * p.Δ
-    n += 1
+    @. v = ρ*v - η*p.Δ
+    @. p.Δ = -d
   end
 end
 
 function rmsprop(p::Param; η::Real = 0.001, ρ::Real = 0.9, ϵ::Real = 1e-8)
-  acc  = zeros(p.x) .+ ϵ
+  acc  = zeros(p.x)
   function ()
-    @. acc = ρ * acc + (1 - ρ) * p.Δ ^ 2
+    @. acc = ρ * acc + (1 - ρ) * p.Δ^2
-    @. p.Δ *= η / √acc
+    @. p.Δ *= η / (√acc + ϵ)
   end
 end
 
 function adagrad(p::Param; η::Real = 0.01, ϵ::Real = 1e-8)
   acc = zeros(p.x) .+ ϵ
   function ()
-    @. acc += p.Δ ^ 2
+    @. acc += p.Δ^2
     @. p.Δ *= η / √acc
   end
 end
 
-function adadelta(p::Param; ρ::Real = 0.95, ϵ::Real = 1e-8)
+function adadelta(p::Param; ρ::Real = 0.9, ϵ::Real = 1e-8)
-  acc = zeros(p.x) .+ ϵ
+  acc = zeros(p.x)
-  Δacc = zeros(p.x) .+ ϵ
+  Δacc = zeros(p.x)
   function ()
-    @. acc = ρ * acc + (1 - ρ) * p.Δ ^ 2
+    @. acc = ρ * acc + (1 - ρ) * p.Δ^2
-    @. p.Δ *= √Δacc / √acc
+    @. p.Δ *= √(Δacc + ϵ) / √(acc + ϵ)
-    @. Δacc = ρ * Δacc + (1 - ρ) * p.Δ ^ 2
+    @. Δacc = ρ * Δacc + (1 - ρ) * p.Δ^2
    end
 end
 
 function adam(p::Param; η::Real = 0.001, β1::Real = 0.9, β2::Real = 0.999, ϵ::Real = 1e-8)
   mt = zeros(p.x)
-  vt = zeros(p.x) .+ ϵ
+  vt = zeros(p.x)
   β1p, β2p = β1, β2
   function ()
     @. mt = β1 * mt + (1 - β1) * p.Δ
-    @. vt = β2 * vt + (1 - β2) * p.Δ ^ 2
+    @. vt = β2 * vt + (1 - β2) * p.Δ^2
-    @. p.Δ = √(1 - β2p) / √(1 - β1p) * mt / √vt * η
+    @. p.Δ =  mt / (1 - β1p) / (√(vt / (1 - β2p)) + ϵ) * η
     β1p *= β1
     β2p *= β2
   end
 end
+
+function amsgrad(p::Param; η::Real = 0.001, β1::Real = 0.9, β2::Real = 0.999, ϵ::Real = 1e-8)
+  mt = zeros(p.x)
+  vt = zeros(p.x) .+ ϵ
+  v̂t = zeros(p.x) .+ ϵ
+  function ()
+    @. mt = β1 * mt + (1 - β1) * p.Δ
+    @. vt = β2 * vt + (1 - β2) * p.Δ ^ 2
+    @. v̂t = max.(v̂t, vt)
+    @. p.Δ = η * mt / √v̂t
+  end
+end
+
+clip(p::Param, thresh::Real) = () -> clamp!(p.Δ, -thresh, thresh)
+
+function expdecay(p::Param, γ::Real)
+  if γ != 0
+    return () -> p.Δ .+= γ .* p.x
+  else
+    return () -> nothing
+  end
+end
+
+function invdecay(p::Param, γ::Real)
+  if γ != 0
+    n = 0
+    return () -> begin
+      p.Δ .*= 1 / (1 + γ * n)
+      n += 1
+    end
+  else
+    return () -> nothing
+  end
+end

src/optimise/train.jl

@@ -1,15 +1,24 @@
 using Juno
-using Flux.Tracker: back!
+using Flux.Tracker: back!, value
 
 runall(f) = f
 runall(fs::AbstractVector) = () -> foreach(call, fs)
 
 """
-    train!(loss, data, opt; cb = () -> ())
+    train!(loss, data, opt)
 
 For each datapoint `d` in `data` computes the gradient of `loss(d...)` through
-backpropagation and calls the optimizer `opt` and the callback `cb`
+backpropagation and calls the optimizer `opt`.
-(i.e. `opt()` and `cb()`).
+
+Takes a callback as keyword argument `cb`. For example, this will print "training"
+every 10 seconds:
+
+```julia
+Flux.train!(loss, data, opt,
+            cb = throttle(() -> println("training"), 10))
+```
+
+The callback can return `:stop` to interrupt the training loop.
 
 Multiple optimisers and callbacks can be passed to `opt` and `cb` as arrays.
 """
@@ -18,10 +27,10 @@ function train!(loss, data, opt; cb = () -> ())
   opt = runall(opt)
   @progress for d in data
     l = loss(d...)
-    isinf(l.data[]) && error("Loss is Inf")
+    isinf(value(l)) && error("Loss is Inf")
-    isnan(l.data[]) && error("Loss is NaN")
+    isnan(value(l)) && error("Loss is NaN")
     back!(l)
     opt()
-    cb()
+    cb() == :stop && break
   end
 end

src/tracker/Tracker.jl

@@ -58,6 +58,7 @@ Base.similar(x::TrackedArray, dims::Union{AbstractUnitRange,Integer}...) =
 
 Base.similar(x::TrackedArray, T::Type) = similar(data(x), T)
 
+# TODO decide if keeping both data and value. The problem is TrackedScalar
 value(x) = x
 value(x::TrackedArray) = data(x)
 value(x::TrackedScalar) = data(x)[]
@@ -69,6 +70,7 @@ Base.:(==)(x::TrackedArray, y::TrackedArray) = value(x) == value(x)
 Base.isless(x::TrackedScalar, y) = isless(value(x), y)
 Base.isless(x, y::TrackedScalar) = isless(x, value(y))
 Base.isless(x::TrackedScalar, y::TrackedScalar) = isless(value(x), value(y))
+Base.isapprox(x::TrackedScalar, y; kws...) = isapprox(x.data[], y; kws...)
 
 Base.show(io::IO, ::Type{TrackedArray{T,N,A}}) where {T,N,A<:AbstractArray{T,N}} =
   print(io, "TrackedArray{…,$A}")
@@ -91,7 +93,7 @@ include("back.jl")
 include("lib.jl")
 include("numeric.jl")
 
-import NNlib.adapt
+import Adapt.adapt
 
 adapt(T, xs::TrackedArray) = TrackedArray(xs.f, adapt(T, xs.data), adapt(T, xs.grad))
 

src/tracker/back.jl

@@ -12,16 +12,17 @@ function scan(x::TrackedArray)
   return
 end
 
-back(c::Call, Δ) = back(c.func, Δ, c.args...)
+back_(f, y, args...) = back(f, args...)
-back(::Call{Void}, Δ) = nothing
+back_(c::Call, y, Δ) = back_(c.func, y, Δ, c.args...)
+back_(::Call{Void}, y, Δ) = nothing
 
 function back(x::TrackedArray, Δ)
   ref = x.ref -= 1
   if isdefined(x, :grad)
     x.grad .+= Δ
-    ref == 0 && back(x.f, x.grad)
+    ref == 0 && back_(x.f, x.data, x.grad)
   else
-    ref == 0 && back(x.f, Δ)
+    ref == 0 && back_(x.f, x.data, Δ)
   end
   return
 end
@@ -35,6 +36,9 @@ end
 
 # Interface methods
 
+# TODO: if an error occurs in `back` the refcounts will be broken
+# and `back` will silently fail to update.
+
 function back!(x::TrackedArray, Δ)
   scan(x)
   back(x, Δ)

src/tracker/lib.jl

@@ -48,6 +48,12 @@ function back(::typeof(vcat), Δ, xs...)
   end
 end
 
+Base.reshape(xs::TrackedArray, dims::Union{Colon,Int64}...) =
+  TrackedArray(Call(reshape, xs, dims...))
+
+back(::typeof(reshape), Δ, xs::TrackedArray, _...) =
+  back(xs, reshape(Δ, size(xs)))
+
 # Reductions
 
 Base.sum(xs::TrackedArray, dim) = TrackedArray(Call(sum, xs, dim))
@@ -62,6 +68,15 @@ Base.findfirst(xs::TrackedArray, args...) = findfirst(xs.data, args...)
 Base.mean(xs::TrackedArray) = TrackedArray(Call(mean, xs), toarray(xs.data, mean(xs.data)))
 Base.mean(xs::TrackedArray, region) = TrackedArray(Call(mean, xs, region))
 
+LinAlg.dot(xs::TrackedVector, ys::TrackedVector) = TrackedArray(Call(dot, xs, ys), toarray(xs.data, dot(data(xs), data(ys))))
+LinAlg.dot(xs::AbstractVector, ys::TrackedVector) = TrackedArray(Call(dot, xs, ys), toarray(xs.data, dot(data(xs), data(ys))))
+LinAlg.dot(xs::TrackedVector, ys::AbstractVector) = TrackedArray(Call(dot, xs, ys), toarray(xs.data, dot(data(xs), data(ys))))
+
+function back(::typeof(dot), Δ, xs, ys)
+  @back(xs, Δ.*ys)
+  @back(ys, Δ.*xs)
+end
+
 # Hacks to get std working
 Base.std(x::TrackedArray; mean = Base.mean(x)) =
   sqrt.(sum((x .- mean).^2) ./ (length(x)-1))
@@ -74,7 +89,7 @@ back(::typeof(mean), Δ, xs::TrackedArray, region) =
 
 # BLAS
 
-for f in :[*, Ac_mul_B].args
+for f in :[*, Ac_mul_B, A_mul_Bc].args
   @eval begin
     import Base.$f
     $f(a::TrackedMatrix, b::TrackedMatrix)  = TrackedArray(Call($f, a, b))
@@ -98,7 +113,12 @@ end
 
 function back(::typeof(Ac_mul_B), Δ, a::AbstractVecOrMat{<:Real}, b::AbstractVecOrMat{<:Real})
   @back(a, A_mul_Bt(Δ, data(b))')
-  @back(b, *(data(a), Δ))
+  @back(b, data(a)*Δ)
+end
+
+function back(::typeof(A_mul_Bc), Δ, a::AbstractVecOrMat{<:Real}, b::AbstractVecOrMat{<:Real})
+  @back(a, Δ * data(b))
+  @back(b, At_mul_B(data(a), Δ)')
 end
 
 # Fast path for matrix-vector
@@ -113,12 +133,36 @@ end
 
 # NNlib
 
-import NNlib: softmax, ∇softmax
+using NNlib
+import NNlib: softmax, ∇softmax, conv2d, pool
 
 softmax(xs::TrackedArray) = TrackedArray(Call(softmax, xs))
 
 back(::typeof(softmax), Δ, xs) = @back(xs, ∇softmax(Δ, data(xs)))
 
+# TODO: can store kwargs efficiently in namedtuples
+_conv2d(x, w, stride, pad) = conv2d(x, w, stride = stride, padding = pad)
+
+conv2d(x::TrackedArray{<:Any,4}, w::TrackedArray{<:Any,4}; stride = 1, padding = 0) =
+  TrackedArray(Call(_conv2d, x, w, stride, padding))
+conv2d(x::AbstractArray{<:Any,4}, w::TrackedArray{<:Any,4}; stride = 1, padding = 0) =
+  TrackedArray(Call(_conv2d, x, w, stride, padding))
+conv2d(x::TrackedArray{<:Any,4}, w::AbstractArray{<:Any,4}; stride = 1, padding = 0) =
+  TrackedArray(Call(_conv2d, x, w, stride, padding))
+
+function back(::typeof(_conv2d), Δ, x, w, stride, pad)
+  @back(x, NNlib.conv2d_grad_x(data(x), data(w), Δ; stride = stride, padding = pad))
+  @back(w, NNlib.conv2d_grad_w(data(x), data(w), Δ; stride = stride, padding = pad))
+end
+
+_pool(x, k, pad, mode) = pool(x, window = k, mode = mode, padding = pad)
+
+pool(x::TrackedArray{<:Any,4}; window = 2, mode = 0, padding = 0) =
+  TrackedArray(Call(_pool, x, window, padding, mode))
+
+back_(::typeof(_pool), y, Δ, x, k, pad, mode) =
+  back(x, NNlib.pool_grad(data(x), y, Δ, window=k, mode=mode, padding=pad))
+
 # Broadcasting
 
 using ForwardDiff: Dual, partials
@@ -134,9 +178,11 @@ dualify(xs::TrackedArray, ps) = map(x -> Dual(x, ps), data(xs))
 
 function tracked_broadcast(f, args::Vararg{Any,N}) where N
   dargs = map((x,i) -> dualify(x, ntuple(j -> i==j, Val{N})), args, ntuple(identity, Val{N}))
+  out = broadcast(f, dargs...)
+  eltype(out) <: Dual || return out
   # TrackedArray(Call(Broadcasted(broadcast(f, dargs...)), args...))
   # Works around a 0.6 type inference issue
-  b = Broadcasted(broadcast(f, dargs...))
+  b = Broadcasted(out)
   TrackedArray(Call(b, args...), b())
 end
 

src/tracker/numeric.jl

@@ -19,4 +19,4 @@ function ngradient(f, xs::AbstractArray...)
   return grads
 end
 
-gradcheck(f, xs...) = all(isapprox.(ngradient(f, xs...), gradient(f, xs...), rtol = 1e-6))
+gradcheck(f, xs...) = all(isapprox.(ngradient(f, xs...), gradient(f, xs...), rtol = 1e-5))

src/utils.jl

@@ -1,8 +1,8 @@
 # Arrays
 
 initn(dims...) = randn(dims...)/100
-
+glorot_uniform(dims...) = (rand(dims...) - 0.5)*sqrt(24.0/(sum(dims)))
-flatten(xs) = reshape(xs, size(xs, 1), :)
+glorot_normal(dims...) = (randn(dims...)*sqrt(2.0/sum(dims)))
 
 unsqueeze(xs, dim) = reshape(xs, (size(xs)[1:dim-1]..., 1, size(xs)[dim:end]...))
 
@@ -93,13 +93,14 @@ but if you'd like to disable the execution on the leading edge, pass
 function throttle(f, timeout; leading=true, trailing=false)
   cooldown = true
   later = nothing
+  result = nothing
 
   function throttled(args...; kwargs...)
     yield()
 
     if cooldown
       if leading
-        f(args...; kwargs...)
+        result = f(args...; kwargs...)
       else
         later = () -> f(args...; kwargs...)
       end
@@ -114,9 +115,28 @@ function throttle(f, timeout; leading=true, trailing=false)
         cooldown = true
       end
     elseif trailing
-      later = () -> f(args...; kwargs...)
+      later = () -> (result = f(args...; kwargs...))
     end
 
-    nothing
+    return result
   end
 end
+
+"""
+    J = jacobian(m,x)
+
+Calculate the output jacobian `J = d/dx m(x)` such that each row `i` of `J` corresponds to the gradient `J[i,:] = ∇ₓ(m(x)[i])`
+"""
+function jacobian(m,x)
+    xp = param(x)
+    y  = m(xp)
+    k  = length(y)
+    n  = length(x)
+    J  = Matrix{eltype(x)}(n,k)
+    for i = 1:k
+        Flux.back!(y[i]) # Populate gradient accumulator
+        J[:,i] = xp.grad
+        xp.grad .*= 0 # Reset gradient accumulator
+    end
+    J'
+end

test/data.jl

@@ -1,3 +1,8 @@
 using Flux.Data
+using Base.Test
 
 @test cmudict()["CATASTROPHE"] == :[K,AH0,T,AE1,S,T,R,AH0,F,IY0].args
+
+@test length(CMUDict.phones()) == 39
+
+@test length(CMUDict.symbols()) == 84

test/layers/normalisation.jl

@@ -26,3 +26,55 @@ using Flux: testmode!
   y = m(x)
   @test count(a->a == 0, y) == 0
 end
+
+@testset "BatchNorm" begin
+  let m = BatchNorm(2), x = param([1 2; 3 4; 5 6]')
+
+    @test m.β.data == [0, 0]  # initβ(2)
+    @test m.γ.data == [1, 1]  # initγ(2)
+    # initial m.σ is 1
+    # initial m.μ is 0
+    @test m.active
+
+    # @test m(x).data ≈ [-1 -1; 0 0; 1 1]'
+    m(x)
+
+    # julia> x
+    #  2×3 Array{Float64,2}:
+    #  1.0  3.0  5.0
+    #  2.0  4.0  6.0
+    #
+    # μ of batch will be
+    #  (1. + 3. + 5.) / 3 = 3
+    #  (2. + 4. + 6.) / 3 = 4
+    #
+    # ∴ update rule with momentum:
+    #  .1 * 3 + 0 = .3
+    #  .1 * 4 + 0 = .4
+    @test m.μ ≈ reshape([0.3, 0.4], 2, 1)
+
+    # julia> .1 .* std(x, 2, corrected=false) .* (3 / 2).+ .9 .* [1., 1.]
+    # 2×1 Array{Float64,2}:
+    #  1.14495
+    #  1.14495
+    @test m.σ ≈ .1 .* std(x.data, 2, corrected=false) .* (3 / 2).+ .9 .* [1., 1.]
+
+    testmode!(m)
+    @test !m.active
+
+    x′ = m(x).data
+    @test x′[1] ≈ (1 - 0.3) / 1.1449489742783179
+  end
+
+  # with activation function
+  let m = BatchNorm(2, λ = σ), x = param([1 2; 3 4; 5 6]')
+    @test m.active
+    m(x)
+
+    testmode!(m)
+    @test !m.active
+
+    x′ = m(x).data
+    @test x′[1] ≈ σ((1 - 0.3) / 1.1449489742783179)
+  end
+end

test/layers/stateless.jl

@@ -0,0 +1,26 @@
+using Flux: onehotbatch, mse, crossentropy
+
+@testset "losses" begin
+  # First, regression-style y's
+  y = [1, 1, 0, 0]
+  y_hat = [.9, .1, .1, .9]
+
+  @testset "mse" begin
+    @test mse(y_hat, y) ≈ (.1^2 + .9^2)/2
+  end
+
+  # Now onehot y's
+  y = onehotbatch([1, 1, 0, 0], 0:1)
+  y_hat = [.1 .9; .9 .1; .9 .1; .1 .9]'
+  y_logloss = 1.203972804325936
+
+  @testset "crossentropy" begin
+    @test crossentropy(y_hat, y) ≈ y_logloss
+  end
+
+  @testset "weighted_crossentropy" begin
+    @test crossentropy(y_hat, y, weight = ones(2)) ≈ y_logloss
+    @test crossentropy(y_hat, y, weight = [.5, .5]) ≈ y_logloss/2
+    @test crossentropy(y_hat, y, weight = [2, .5]) ≈ 1.5049660054074199
+  end
+end

test/optimise.jl

@@ -0,0 +1,29 @@
+using Flux.Optimise
+using Flux.Tracker
+
+@testset "Optimise" begin
+  w = randn(10, 10)
+  for Opt in [SGD, Nesterov, Momentum, ADAM, RMSProp, ps -> ADAGrad(ps, 0.1), ADADelta, AMSGrad]
+    w′ = param(randn(10, 10))
+    loss(x) = Flux.mse(w*x, w′*x)
+    opt = Opt([w′])
+    for t=1:10^5
+      l = loss(rand(10))
+      back!(l)
+      opt()
+    end
+    @test Flux.mse(w, w′) < 0.01
+  end
+end
+
+@testset "Training Loop" begin
+  i = 0
+  l = param(1)
+
+  Flux.train!(() -> (sleep(0.1); i += 1; l),
+              Iterators.repeated((), 100),
+              ()->(),
+              cb = Flux.throttle(() -> (i > 3 && :stop), 1))
+
+  @test 3 < i < 50
+end

test/runtests.jl

@@ -5,5 +5,8 @@ using Flux, Base.Test
 include("utils.jl")
 include("tracker.jl")
 include("layers/normalisation.jl")
+include("layers/stateless.jl")
+include("optimise.jl")
+include("data.jl")
 
 end

test/tracker.jl

@@ -1,5 +1,6 @@
 using Flux.Tracker, Base.Test, NNlib
 using Flux.Tracker: gradcheck
+using NNlib
 
 gradtest(f, xs::AbstractArray...) = gradcheck((xs...) -> sum(f(xs...)), xs...)
 gradtest(f, dims...) = gradtest(f, rand.(dims)...)
@@ -10,6 +11,7 @@ gradtest(f, dims...) = gradtest(f, rand.(dims)...)
 @test gradtest((x, W, b) -> σ.(W*x .+ b), (5,3), (2,5), 2)
 
 @test gradtest((w, x) -> w'*x, randn(10, 2), randn(10))
+@test gradtest((w, x) -> w*x', randn(5,5), randn(5,5))
 
 @test gradtest(x -> sin.(sum(x, (2, 3))), (3,4,5))
 
@@ -37,9 +39,17 @@ end
 @test gradtest(x -> std(x), rand(5,5))
 @test gradtest(x -> std(x, 1), rand(5,5))
 
+@test gradtest((x, y) -> x .* y, rand(5), rand(5))
+
 @test gradtest(rand(5)) do x
   y = x.^2
   2y + x
 end
 
+@test gradtest(conv2d, rand(10, 10, 3, 2), randn(2, 2, 3, 2))
+@test gradtest(x -> maxpool2d(x, 2), rand(10, 10, 3, 2))
+@test gradtest(x -> avgpool2d(x, 2), rand(10, 10, 3, 2))
+
+@test (param([1,2,3]) .< 2) == [true, false, false]
+
 end #testset

test/utils.jl

@@ -1,4 +1,4 @@
-using Flux: throttle
+using Flux: throttle, initn, glorot_uniform, glorot_normal, jacobian
 
 @testset "Throttle" begin
   @testset "default behaviour" begin
@@ -47,3 +47,35 @@ using Flux: throttle
     @test a == [1, 3]
   end
 end
+
+@testset "Jacobian" begin
+  A = param(randn(2,2))
+  x = randn(2)
+  m(x) = A*x
+  y = m(x)
+  J = jacobian(m,x)
+  @test J ≈ A.data
+end
+
+@testset "Initialization" begin
+  # Set random seed so that these tests don't fail randomly
+  srand(0)
+  # initn() should yield a kernel with stddev ~= 1e-2
+  v = initn(10, 10)
+  @test std(v) > 0.9*1e-2
+  @test std(v) < 1.1*1e-2
+
+  # glorot_uniform should yield a kernel with stddev ~= sqrt(6/(n_in + n_out)),
+  # and glorot_normal should yield a kernel with stddev != 2/(n_in _ n_out)
+  for (n_in, n_out) in [(100, 100), (100, 400)]
+    v = glorot_uniform(n_in, n_out)
+    @test minimum(v) > -1.1*sqrt(6/(n_in + n_out))
+    @test minimum(v) < -0.9*sqrt(6/(n_in + n_out))
+    @test maximum(v) >  0.9*sqrt(6/(n_in + n_out))
+    @test maximum(v) <  1.1*sqrt(6/(n_in + n_out))
+
+    v = glorot_normal(n_in, n_out)
+    @test std(v) > 0.9*sqrt(2/(n_in + n_out))
+    @test std(v) < 1.1*sqrt(2/(n_in + n_out))
+  end
+end