Avoiding Type Piracy in Lux

.github/workflows/ci.yml

@@ -18,8 +18,8 @@ jobs:
           - '1.6'
           - '1.8'
           - '1.9'
+          - '1.10'
           - '1' # Leave this line unchanged. '1' will automatically expand to the latest stable 1.x release of Julia.
-          - '1.10.0-beta3'
         os:
           - ubuntu-latest
         arch:

Project.toml

@@ -1,7 +1,7 @@
 name = "ComponentArrays"
 uuid = "b0b7db55-cfe3-40fc-9ded-d10e2dbeff66"
 authors = ["Jonnie Diegelman <[email protected]>"]
-version = "0.15.10"
+version = "0.15.11"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
@@ -17,20 +17,24 @@ StaticArraysCore = "1e83bf80-4336-4d27-bf5d-d5a4f845583c"
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 ConstructionBase = "187b0558-2788-49d3-abe0-74a17ed4e7c9"
 GPUArrays = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
+Optimisers = "3bd65402-5787-11e9-1adc-39752487f4e2"
 RecursiveArrayTools = "731186ca-8d62-57ce-b412-fbd966d074cd"
 ReverseDiff = "37e2e3b7-166d-5795-8a7a-e32c996b4267"
 SciMLBase = "0bca4576-84f4-4d90-8ffe-ffa030f20462"
 Tracker = "9f7883ad-71c0-57eb-9f7f-b5c9e6d3789c"
+TruncatedStacktraces = "781d530d-4396-4725-bb49-402e4bee1e77"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [extensions]
 ComponentArraysAdaptExt = "Adapt"
 ComponentArraysConstructionBaseExt = "ConstructionBase"
 ComponentArraysGPUArraysExt = "GPUArrays"
+ComponentArraysOptimisersExt = "Optimisers"
 ComponentArraysRecursiveArrayToolsExt = "RecursiveArrayTools"
 ComponentArraysReverseDiffExt = "ReverseDiff"
 ComponentArraysSciMLBaseExt = "SciMLBase"
 ComponentArraysTrackerExt = "Tracker"
+ComponentArraysTruncatedStacktracesExt = "TruncatedStacktraces"
 ComponentArraysZygoteExt = "Zygote"
 
 [compat]
@@ -41,13 +45,17 @@ ConstructionBase = "1"
 ForwardDiff = "0.10"
 Functors = "0.4.4"
 GPUArrays = "8, 9, 10"
+LinearAlgebra = "1"
+Optimisers = "0.3"
 PackageExtensionCompat = "1"
 RecursiveArrayTools = "2, 3"
 ReverseDiff = "1"
 SciMLBase = "1, 2"
-StaticArraysCore = "1"
 StaticArrayInterface = "1"
+StaticArraysCore = "1"
+Test = "1"
 Tracker = "0.2"
+TruncatedStacktraces = "1.4"
 Zygote = "0.6"
 julia = "1.6"
 
@@ -56,9 +64,11 @@ Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 ConstructionBase = "187b0558-2788-49d3-abe0-74a17ed4e7c9"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 GPUArrays = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
+Optimisers = "3bd65402-5787-11e9-1adc-39752487f4e2"
 RecursiveArrayTools = "731186ca-8d62-57ce-b412-fbd966d074cd"
 ReverseDiff = "37e2e3b7-166d-5795-8a7a-e32c996b4267"
 SciMLBase = "0bca4576-84f4-4d90-8ffe-ffa030f20462"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 Tracker = "9f7883ad-71c0-57eb-9f7f-b5c9e6d3789c"
+TruncatedStacktraces = "781d530d-4396-4725-bb49-402e4bee1e77"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"

ext/ComponentArraysOptimisersExt.jl

@@ -0,0 +1,21 @@
+module ComponentArraysOptimisersExt
+
+using ComponentArrays, Optimisers
+
+# Optimisers can handle componentarrays by default, but we can vectorize the entire
+# operation here instead of doing multiple smaller operations
+Optimisers.setup(opt::AbstractRule, ps::ComponentArray) = Optimisers.setup(opt, getdata(ps))
+
+function Optimisers.update(tree, ps::ComponentArray, gs::ComponentArray)
+    gs_flat = ComponentArrays.__value(getdata(gs)) # Safety against ReverseDiff
+    tree, ps_new = Optimisers.update(tree, getdata(ps), gs_flat)
+    return tree, ComponentArray(ps_new, getaxes(ps))
+end
+
+function Optimisers.update!(tree::Optimisers.Leaf, ps::ComponentArray, gs::ComponentArray)
+    gs_flat = ComponentArrays.__value(getdata(gs)) # Safety against ReverseDiff
+    tree, ps_new = Optimisers.update!(tree, getdata(ps), gs_flat)
+    return tree, ComponentArray(ps_new, getaxes(ps))
+end
+
+end

ext/ComponentArraysReverseDiffExt.jl

@@ -2,7 +2,7 @@ module ComponentArraysReverseDiffExt
 
 using ComponentArrays, ReverseDiff
 
-const TrackedComponentArray{V, D, N, DA, A, Ax} = ReverseDiff.TrackedArray{V,D,N,ComponentArray{V,N,A,Ax},DA}
+const TrackedComponentArray{V,D,N,DA,A,Ax} = ReverseDiff.TrackedArray{V,D,N,ComponentArray{V,N,A,Ax},DA}
 
 maybe_tracked_array(val::AbstractArray, der, tape, inds, origin) = ReverseDiff.TrackedArray(val, der, tape)
 function maybe_tracked_array(val::Real, der, tape, inds, origin::AbstractVector)
@@ -12,10 +12,10 @@ function maybe_tracked_array(val::Real, der, tape, inds, origin::AbstractVector)
 end
 
 for f in [:getindex, :view]
-    @eval function Base.$f(tca::TrackedComponentArray, inds::Union{Symbol, Val}...)
-            val = $f(ReverseDiff.value(tca), inds...)
-            der = Base.maybeview(ReverseDiff.deriv(tca), inds...)
-            t = ReverseDiff.tape(tca)
+    @eval function Base.$f(tca::TrackedComponentArray, inds::Union{Symbol,Val}...)
+        val = $f(ReverseDiff.value(tca), inds...)
+        der = Base.maybeview(ReverseDiff.deriv(tca), inds...)
+        t = ReverseDiff.tape(tca)
         return maybe_tracked_array(val, der, t, inds, tca)
     end
 end
@@ -31,4 +31,17 @@ function Base.getproperty(tca::TrackedComponentArray, s::Symbol)
     end
 end
 
+function Base.propertynames(::TrackedComponentArray{V,D,N,DA,A,Tuple{Ax}}) where {V,D,N,DA,A,Ax<:ComponentArrays.AbstractAxis}
+    return propertynames(ComponentArrays.indexmap(Ax))
+end
+
+function Base.NamedTuple(tca::TrackedComponentArray)
+    props = propertynames(tca)
+    return NamedTuple{props}(getproperty(tca, p) for p in props)
+end
+
+@inline ComponentArrays.__value(x::AbstractArray{<:ReverseDiff.TrackedReal}) = ReverseDiff.value.(x)
+@inline ComponentArrays.__value(x::ReverseDiff.TrackedArray) = ReverseDiff.value(x)
+@inline ComponentArrays.__value(x::TrackedComponentArray) = ComponentArray(ComponentArrays.__value(getdata(x)), getaxes(x))
+
 end

ext/ComponentArraysTruncatedStacktracesExt.jl

@@ -0,0 +1,8 @@
+module ComponentArraysTruncatedStacktracesExt
+
+using ComponentArrays
+import TruncatedStacktraces: @truncate_stacktrace
+
+@truncate_stacktrace ComponentArray 1
+
+end

src/ComponentArrays.jl

@@ -52,6 +52,8 @@ include("compat/chainrulescore.jl")
 include("compat/static_arrays.jl")
 export @static_unpack
 
+include("compat/functors.jl")
+
 import PackageExtensionCompat: @require_extensions
 function __init__()
     @require_extensions

src/compat/chainrulescore.jl

@@ -41,7 +41,26 @@ end
 # Prevent double projection
 (p::ChainRulesCore.ProjectTo{ComponentArray})(dx::ComponentArray) = dx
 
-function (p::ChainRulesCore.ProjectTo{ComponentArray})(t::ChainRulesCore.Tangent{A, <:NamedTuple}) where {A}
+function (p::ChainRulesCore.ProjectTo{ComponentArray})(t::ChainRulesCore.Tangent{A,<:NamedTuple}) where {A}
     nt = Functors.fmap(ChainRulesCore.backing, ChainRulesCore.backing(t))
     return ComponentArray(nt)
 end
+
+function ChainRulesCore.rrule(::Type{CA}, nt::NamedTuple) where {CA<:ComponentArray}
+    y = CA(nt)
+
+    function ∇NamedTupleToComponentArray(Δ::AbstractArray)
+        if length(Δ) == length(y)
+            return ∇NamedTupleToComponentArray(ComponentArray(vec(Δ), getaxes(y)))
+        end
+        error("Got pullback input of shape $(size(Δ)) & type $(typeof(Δ)) for output " *
+              "of shape $(size(y)) & type $(typeof(y))")
+        return nothing
+    end
+
+    function ∇NamedTupleToComponentArray(Δ::ComponentArray)
+        return ChainRulesCore.NoTangent(), NamedTuple(Δ)
+    end
+
+    return y, ∇NamedTupleToComponentArray
+end

src/compat/functors.jl

@@ -0,0 +1 @@
+Functors.functor(::Type{<:ComponentVector}, c) = NamedTuple(c), ComponentVector

src/utils.jl

@@ -50,3 +50,5 @@ recursive_eltype(x::AbstractArray{<:Any}) = isempty(x) ? Base.Bottom : mapreduce
 recursive_eltype(x::Dict) = isempty(x) ? Base.Bottom : mapreduce(recursive_eltype, promote_type, values(x))
 recursive_eltype(::AbstractArray{T,N}) where {T<:Number, N} = T
 recursive_eltype(x) = typeof(x)
+
+@inline __value(x) = x

test/Project.toml

@@ -9,9 +9,11 @@ JLArrays = "27aeb0d3-9eb9-45fb-866b-73c2ecf80fcb"
 LabelledArrays = "2ee39098-c373-598a-b85f-a56591580800"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 OffsetArrays = "6fe1bfb0-de20-5000-8ca7-80f57d26f881"
+Optimisers = "3bd65402-5787-11e9-1adc-39752487f4e2"
 ReverseDiff = "37e2e3b7-166d-5795-8a7a-e32c996b4267"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 Tracker = "9f7883ad-71c0-57eb-9f7f-b5c9e6d3789c"
+TruncatedStacktraces = "781d530d-4396-4725-bb49-402e4bee1e77"
 Unitful = "1986cc42-f94f-5a68-af5c-568840ba703d"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"

test/autodiff_tests.jl

@@ -1,5 +1,5 @@
 import FiniteDiff, ForwardDiff, ReverseDiff, Tracker, Zygote
-
+using Optimisers
 using Test
 
 F(a, x) = sum(abs2, a) * x^3
@@ -38,6 +38,22 @@ truth = ComponentArray(a = [32, 48], x = 156)
     @test out isa Vector{<:ForwardDiff.Dual}
 end
 
+@testset "Optimisers Update" begin
+    ca_ = deepcopy(ca)
+    opt_st = Optimisers.setup(Adam(0.01), ca_)
+    gs_zyg = only(Zygote.gradient(F_idx_val, ca_))
+    @test !(last(Optimisers.update(opt_st, ca_, gs_zyg)) ≈ ca)
+    Optimisers.update!(opt_st, ca_, gs_zyg)
+    @test !(ca_ ≈ ca)
+
+    ca_ = deepcopy(ca)
+    opt_st = Optimisers.setup(Adam(0.01), ca_)
+    gs_rdiff = ReverseDiff.gradient(F_idx_val, ca_)
+    @test !(last(Optimisers.update(opt_st, ca_, gs_rdiff)) ≈ ca)
+    Optimisers.update!(opt_st, ca_, gs_rdiff)
+    @test !(ca_ ≈ ca)
+end
+
 @testset "Projection" begin
     gs_ca = Zygote.gradient(sum, ca)[1]
 
@@ -76,18 +92,28 @@ end
     @test ∂r ≈ ∂r_ca
 end
 
+function F_prop(x)
+    @assert propertynames(x) == (:x, :y)
+    return sum(abs2, x.x .- x.y)
+end
+
+@testset "Preserve Properties" begin
+    x = ComponentArray(; x = [1.0, 5.0], y = [3.0, 4.0])
 
-# # This is commented out because the gradient operation itself is broken due to Zygote's inability
-# # to support mutation and ComponentArray's use of mutation for construction from a NamedTuple.
-# # It would be nice to support this eventually, so I'll just leave this commented (because @test_broken
-# # wouldn't work here because the error happens before the test)
-# @testset "Issues" begin
-#     function mysum(x::AbstractVector)
-#         y = ComponentVector(x=x)
-#         return sum(y)
-#     end
+    gs_z = only(Zygote.gradient(F_prop, x))
+    gs_rdiff = ReverseDiff.gradient(F_prop, x)
 
-#     Δ = Zygote.gradient(mysum, rand(10))
+    @test gs_z ≈ gs_rdiff
+end
+
+@testset "Issues" begin
+    function mysum(x::AbstractVector)
+        y = ComponentVector(x=x)
+        z = ComponentVector(; z = x .^ 2)
+        return sum(y) + sum(abs2, z)
+    end
 
-#     @test Δ isa Vector{Float64}
-# end
+    Δ = only(Zygote.gradient(mysum, rand(10)))
+
+    @test Δ isa AbstractVector{Float64}
+end

test/runtests.jl

@@ -8,6 +8,8 @@ using StaticArrays
 using OffsetArrays
 using Test
 using Unitful
+using Functors
+import TruncatedStacktraces # This is loaded just to trigger the extension package
 
 
 ## Test setup
@@ -690,6 +692,14 @@ end
     @test_throws ArgumentError axpby!(2, x, 3, y)
 end
 
+@testset "Functors" begin
+    for carray in (ca, ca_Float32, ca_MVector, ca_SVector, ca_composed, ca2, caa)
+        θ, re = Functors.functor(carray)
+        @test θ isa NamedTuple
+        @test re(θ) == carray
+    end
+end
+
 @testset "Autodiff" begin
     include("autodiff_tests.jl")
 end