Merge branch 'master' into sf/weighted_crossentropy

docs/src/models/layers.md

@@ -37,6 +37,7 @@ These layers don't affect the structure of the network but may improve training
 
 ```@docs
 Flux.testmode!
+BatchNorm
 Dropout
 LayerNorm
 ```

src/Flux.jl

@@ -7,8 +7,9 @@ module Flux
 using Juno, Requires
 using Lazy: @forward
 
-export Chain, Dense, RNN, LSTM, Dropout, LayerNorm,
+export Chain, Dense, RNN, LSTM,
-  SGD, ADAM, Momentum, Nesterov,
+  Dropout, LayerNorm, BatchNorm,
+  SGD, ADAM, Momentum, Nesterov, AMSGrad,
   param, params, mapleaves
 
 using NNlib

src/data/cmudict.jl

@@ -33,7 +33,7 @@ function rawdict()
        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/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/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
 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/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}")

src/tracker/lib.jl

@@ -58,6 +58,15 @@ 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))
@@ -70,7 +79,7 @@ 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))
@@ -94,7 +103,12 @@ 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

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/optimise.jl

@@ -0,0 +1,17 @@
+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

test/runtests.jl

@@ -6,5 +6,6 @@ include("utils.jl")
 include("tracker.jl")
 include("layers/normalisation.jl")
 include("layers/stateless.jl")
+include("optimise.jl")
 
 end

test/tracker.jl

@@ -10,6 +10,7 @@ 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))
 
@@ -37,6 +38,8 @@ 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