trainables

Author: CarloLucibello

docs/src/api.md

@@ -58,6 +58,7 @@ Such restrictions are also obeyed by this function for flattening a model:
 ```@docs
 Optimisers.destructure
 Optimisers.Restructure
+Optimisers.trainables
 ```
 
 ## Rule Definition

src/Optimisers.jl

@@ -11,6 +11,9 @@ include("adjust.jl")
 include("destructure.jl")
 export destructure
 
+include("trainables.jl")
+export trainables
+
 include("rules.jl")
 export Descent, Adam, Momentum, Nesterov, Rprop, RMSProp,
        AdaGrad, AdaMax, AdaDelta, AMSGrad, NAdam, AdamW, RAdam, OAdam, AdaBelief,

src/destructure.jl

@@ -66,7 +66,7 @@ function _flatten(x)
   isnumeric(x) && return vcat(_vec(x)), 0, length(x)  # trivial case
   arrays = AbstractVector[]
   len = Ref(0)
-  off = fmap(x; exclude = isnumeric, walk = _TrainableStructWalk()) do y
+  off = fmap(x; exclude = isnumeric, walk = TrainableStructWalk()) do y
     push!(arrays, _vec(y))
     o = len[]
     len[] = o + length(y)
@@ -76,9 +76,9 @@ function _flatten(x)
   return reduce(vcat, arrays), off, len[]
 end
 
-struct _TrainableStructWalk <: AbstractWalk end
+struct TrainableStructWalk <: AbstractWalk end
 
-(::_TrainableStructWalk)(recurse, x) = map(recurse, _trainable(x))
+(::TrainableStructWalk)(recurse, x) = map(recurse, _trainable(x))
 
 _vec(x::Number) = LinRange(x,x,1)
 _vec(x::AbstractArray) = vec(x)

src/trainables.jl

@@ -1,115 +1,52 @@
-using BenchmarkTools
-using Optimisers
-using Functors
-using Zygote, Flux
-using ChainRulesCore
 
-function trainables1(x)
+"""
+    trainables(x)
+
+Return an iterable over all the trainable parameters in `x`, that is all the numerical
+arrays (see [`isnumeric`](@ref Optimisers.isnumeric)) which are reachable through [`trainable`](@ref Optimisers.trainable).
+
+Parameters appearing multiple times in the model will be present only once in the output.
+
+See also [`destructure`](@ref).
+
+# Examples
+
+```jldoctest
+julia> struct MyLayer
+         w
+         b
+       end
+
+julia> Functors.@functor MyLayer
+
+julia> Optimisers.trainable(x::MyLayer) = (; w = x.w,) # only w is trainable in this example
+
+julia> x = MyLayer([1.0,2.0,3.0], [4.0,5.0,6.0]);
+
+julia> trainables(x)
+1-element Vector{AbstractArray}:
+ [1.0, 2.0, 3.0]
+"""
+function trainables(x)
     arrays = AbstractArray[]
     exclude(x) = Optimisers.isnumeric(x)
-    fmap(x; exclude, walk = Optimisers._TrainableStructWalk()) do y
+    fmap(x; exclude, walk = Optimisers.TrainableStructWalk()) do y
         push!(arrays, y)
         return y
     end
     return arrays
 end
 
-function ∇trainables1(x, Δ)
+function ∇trainables(x, Δ)
     exclude(x) = Optimisers.isnumeric(x)
     i = 0
-    return fmapstructure(x; exclude, walk = Optimisers._TrainableStructWalk()) do _
+    return fmapstructure(x; exclude, walk = Optimisers.TrainableStructWalk()) do _
                 return Δ[i+=1]
            end
 end
 
-
-function ChainRulesCore.rrule(::typeof(trainables1), x)
-    y = trainables1(x)
-    trainables_back(Δ) = (NoTangent(), ∇trainables1(x, unthunk(Δ)))
+function ChainRulesCore.rrule(::typeof(trainables), x)
+    y = trainables(x)
+    trainables_back(Δ) = (NoTangent(), ∇trainables(x, unthunk(Δ)))
     return y, trainables_back
 end
-
-############
-
-using Functors: AbstractWalk, _map, _values, execute, ExcludeWalk
-
-struct TrainableWalk2 <: AbstractWalk end
-
-function (walk::TrainableWalk2)(recurse, x, ys...)
-    x_children = Optimisers.trainable(x)
-    ys_children = map(Optimisers.trainable, ys)
-    res = map(recurse, x_children, ys_children...)
-    return reduce(vcat, values(res),init=[])
-end
-
-function trainables2(x)
-    exclude(x) = Optimisers.isnumeric(x) && Functors.isleaf(x)
-    return execute(ExcludeWalk(TrainableWalk2(), x ->[x], exclude), x)
-end
-
-
-struct TrainableWalk3 <: AbstractWalk end
-
-function (walk::TrainableWalk3)(recurse, x, ys...)
-    x_children = Optimisers.trainable(x)
-    ys_children = map(Optimisers.trainable, ys)
-    res = map(recurse, x_children, ys_children...)
-    return vcat(values(res)...)
-end
-
-function trainables3(x)
-    exclude(x) = Optimisers.isnumeric(x)
-    return execute(ExcludeWalk(TrainableWalk3(), x ->[x], exclude), x)
-end
-
-
-function floss(ps)
-    sum([sum(abs2, p) for p in ps])
-end
-
-using Flux
-
-function perf()
-    m = Chain(Dense(128 => 128, relu), 
-              Dense(128 => 128, relu),
-              BatchNorm(128),
-              x -> x^2,
-              Dense(128 => 128, relu), 
-              Dense(128 => 128, relu))
-              
-    println("trainables1")
-    @btime floss(trainables1($m))
-    println("trainables2")
-    @btime floss(trainables2($m))
-    println("trainables3")
-    @btime floss(trainables3($m))
-    println()
-
-    println("gradient trainables1")
-    @btime gradient(m -> floss(trainables1(m)), $m)
-    println("gradient trainables2")
-    @btime gradient(m -> floss(trainables2(m)), $m)
-    println("gradient trainables3")
-    @btime gradient(m -> floss(trainables3(m)), $m)
-
-    nothing
-end
-
-Zygote.refresh()
-perf()
-
-
-m = Chain(Dense(128 => 128, relu), 
-              Dense(128 => 128, relu),
-              BatchNorm(128),
-              x -> x^2,
-              Dense(128 => 128, relu), 
-              Dense(128 => 128, relu))
-              
-floss(trainables1(m))
-g1 = gradient(m -> floss(trainables1(m)), m)[1]
-g2 = gradient(m -> floss(trainables2(m)), m)[1]
-@test g1.layers[1].weight ≈ g2.layers[1].weight
-@test g1.layers[1].weight ≈ g2.layers[1].weight
-@test g1.layers[3].μ === nothing
-@test g2.layers[3].μ === nothing

test/runtests.jl

@@ -10,7 +10,7 @@ Random.seed!(1)
 
 struct Foo; x; y; end
 Functors.@functor Foo
-Optimisers.trainable(x::Foo) = (x.y, x.x)
+Optimisers.trainable(x::Foo) = (; x.y, x.x)
 
 struct TwoThirds a; b; c; end
 Functors.@functor TwoThirds (a, c)
@@ -539,6 +539,9 @@ end
   @testset verbose=true "Destructure" begin
     include("destructure.jl")
   end
+  @testset verbose=true "Trainables" begin
+    include("trainables.jl")
+  end
   @testset verbose=true "Optimisation Rules" begin
     include("rules.jl")
   end

test/trainables.jl

@@ -0,0 +1,99 @@
+
+m1 = collect(1:3.0)
+m2 = (collect(1:3.0), collect(4:6.0))
+m3 = (x = m1, y = sin, z = collect(4:6.0))
+
+m4 = (x = m1, y = m1, z = collect(4:6.0))  # tied
+m5 = (a = (m3, true), b = (m1, false), c = (m4, true))
+m6 = (a = m1, b = [4.0 + im], c = m1)
+
+m7 = TwoThirds((sin, collect(1:3.0)), (cos, collect(4:6.0)), (tan, collect(7:9.0)))
+m8 = [Foo(m1, m1), (a = true, b = Foo([4.0], false), c = ()), [[5.0]]]
+
+mat = Float32[4 6; 5 7]
+m9 = (a = m1, b = mat, c = [mat, m1])
+
+@testset "trainables" begin
+  ps = trainables(m1)
+  @test ps isa Vector
+  @test length(ps) == 1
+  @test ps[1] == m1
+
+  ps = trainables(m2)
+  @test ps isa Vector
+  @test length(ps) == 2
+  @test ps[1] == m2[1]
+  @test ps[2] == m2[2]
+
+  ps = trainables(m3)
+  @test length(ps) == 2
+  @test ps[1] == 1:3
+  @test ps[2] == 4:6
+
+  ps = trainables(m4)
+  @test length(ps) == 2
+  @test ps[1] == 1:3
+  @test ps[2] == 4:6
+  
+  ps = trainables(m5)
+  @test length(ps) == 3
+  @test ps[1] == 1:3
+  @test ps[2] == 4:6
+  @test ps[3] == 4:6
+
+  ps = trainables(m6)
+  @test length(ps) == 2
+  @test ps[1] == 1:3
+  @test ps[2] == ComplexF64[4.0 + 1.0im]
+
+  ps = trainables(m7)
+  @test length(ps) == 1
+  @test ps[1] == [1.0, 2.0, 3.0]
+
+  ps = trainables(m8)
+  @test length(ps) == 3
+  @test ps[1] == 1:3
+  @test ps[2] == [4.0]
+  @test ps[3] == [5.0]
+
+  ps = trainables(m9)
+  @test length(ps) == 2
+  @test ps[1] == 1:3
+  @test ps[2] == mat
+end
+
+@testset "gradient" begin
+  loss(m) = sum([sum(abs2, p) for p in trainables(m)])
+  g = gradient(loss, m1)[1]
+  @test g == [2.0, 4.0, 6.0]
+
+  g = gradient(loss, m2)[1]
+  @test g == ([2.0, 4.0, 6.0], [8.0, 10.0, 12.0])
+
+  g = gradient(loss, m3)[1]
+  @test g.x == [2.0, 4.0, 6.0]
+  @test g.y === nothing
+  @test g.z == [8.0, 10.0, 12.0]
+
+  g = gradient(loss, m4)[1]
+  @test g == (x = [2.0, 4.0, 6.0], y = [2.0, 4.0, 6.0], z = [8.0, 10.0, 12.0])
+  g.x === g.y # shared gradient for shared weights
+
+  g = gradient(loss, m5)[1]
+  @test g == (a = ((x = [2.0, 4.0, 6.0], y = nothing, z = [8.0, 10.0, 12.0]), nothing), b = ([2.0, 4.0, 6.0], nothing), c = ((x = [2.0, 4.0, 6.0], y = [2.0, 4.0, 6.0], z = [8.0, 10.0, 12.0]), nothing))
+  
+  g = gradient(loss, m6)[1]
+  @test g = (a = [2.0, 4.0, 6.0], b = ComplexF64[8.0 + 2.0im], c = [2.0, 4.0, 6.0])
+  
+  g = gradient(loss, m7)[1]
+  @test g == (a = (nothing, [2.0, 4.0, 6.0]), b = nothing, c = nothing)
+
+  g = gradient(loss, m8)[1]
+  @test g[1] == (x = [2.0, 4.0, 6.0], y = [2.0, 4.0, 6.0])
+  @test g[2] == (a = nothing, b = (x = [8.0], y = nothing), c = nothing)
+  @test g[3] == [[10.0]]
+
+  g = gradient(loss, m9)[1]
+  @test g == (a = [2.0, 4.0, 6.0], b = Float32[8.0 12.0; 10.0 14.0], c = Array[Float32[8.0 12.0; 10.0 14.0], [2.0, 4.0, 6.0]])
+end
+