Fix zygote compat

Author: tansongchen

Project.toml

@@ -6,23 +6,22 @@ version = "0.2.4"
 [deps]
 ChainRules = "082447d4-558c-5d27-93f4-14fc19e9eca2"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
-ChainRulesOverloadGeneration = "f51149dc-2911-5acf-81fc-2076a2a81d4f"
 SymbolicUtils = "d1185830-fcd6-423d-90d6-eec64667417b"
 Symbolics = "0c5d862f-8b57-4792-8d23-62f2024744c7"
-Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [weakdeps]
 NNlib = "872c559c-99b0-510c-b3b7-b6c96a88d5cd"
 SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
+Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [extensions]
 TaylorDiffNNlibExt = ["NNlib"]
 TaylorDiffSFExt = ["SpecialFunctions"]
+TaylorDiffZygoteExt = ["Zygote"]
 
 [compat]
 ChainRules = "1"
 ChainRulesCore = "1"
-ChainRulesOverloadGeneration = "0.1"
 NNlib = "0.9"
 SpecialFunctions = "2"
 SymbolicUtils = "2, 3"

ext/TaylorDiffZygoteExt.jl

@@ -0,0 +1,22 @@
+module TaylorDiffZygoteExt
+
+using TaylorDiff
+import Zygote: @adjoint, Numeric, _dual_safearg, ZygoteRuleConfig
+using ChainRulesCore: @opt_out
+
+# Zygote can't infer this constructor function
+# defining rrule for this doesn't seem to work for Zygote
+# so need to use @adjoint
+@adjoint TaylorScalar{T, N}(t::TaylorScalar{T, M}) where {T, N, M} = TaylorScalar{T, N}(t),
+x̄ -> (TaylorScalar{T, M}(x̄),)
+
+# Zygote will try to use ForwardDiff to compute broadcast functions
+# However, TaylorScalar is not dual safe, so we opt out of this
+_dual_safearg(::Numeric{<:TaylorScalar}) = false
+
+# Zygote has a rule for literal power, need to opt out of this
+@opt_out rrule(
+    ::ZygoteRuleConfig, ::typeof(Base.literal_pow), ::typeof(^), x::TaylorScalar, ::Val{p}
+) where {p}
+
+end

src/chainrules.jl

@@ -1,6 +1,5 @@
-import ChainRulesCore: rrule, RuleConfig, ProjectTo, backing
+import ChainRulesCore: rrule, RuleConfig, ProjectTo, backing, @opt_out
 using Base.Broadcast: broadcasted
-import Zygote: @adjoint, accum_sum, unbroadcast, Numeric, ∇getindex, _project
 
 function contract(a::TaylorScalar{T, N}, b::TaylorScalar{S, N}) where {T, S, N}
     mapreduce(*, +, value(a), value(b))
@@ -31,7 +30,7 @@ function rrule(::typeof(extract_derivative), t::TaylorScalar{T, N},
 end
 
 function rrule(::typeof(*), A::AbstractMatrix{S},
-        t::AbstractVector{TaylorScalar{T, N}}) where {N, S, T}
+        t::AbstractVector{TaylorScalar{T, N}}) where {N, S <: Real, T <: Real}
     project_A = ProjectTo(A)
     function gemv_pullback(x̄)
         x̂ = reinterpret(reshape, T, x̄)
@@ -42,7 +41,7 @@ function rrule(::typeof(*), A::AbstractMatrix{S},
 end
 
 function rrule(::typeof(*), A::AbstractMatrix{S},
-        B::AbstractMatrix{TaylorScalar{T, N}}) where {N, S, T}
+        B::AbstractMatrix{TaylorScalar{T, N}}) where {N, S <: Real, T <: Real}
     project_A = ProjectTo(A)
     project_B = ProjectTo(B)
     function gemm_pullback(x̄)
@@ -54,88 +53,36 @@ function rrule(::typeof(*), A::AbstractMatrix{S},
     return A * B, gemm_pullback
 end
 
-@adjoint function +(t::Vector{TaylorScalar{T, N}}, v::Vector{T}) where {N, T}
-    project_v = ProjectTo(v)
-    t + v, x̄ -> (x̄, project_v(x̄))
-end
-
-@adjoint function +(v::Vector{T}, t::Vector{TaylorScalar{T, N}}) where {N, T}
-    project_v = ProjectTo(v)
-    v + t, x̄ -> (project_v(x̄), x̄)
-end
-
-(project::ProjectTo{T})(dx::TaylorScalar{T, N}) where {N, T} = primal(dx)
-
-# Not-a-number patches
-
-ProjectTo(::T) where {T <: TaylorScalar} = ProjectTo{T}()
-(p::ProjectTo{T})(x::T) where {T <: TaylorScalar} = x
-function ProjectTo(x::AbstractArray{T}) where {T <: TaylorScalar}
-    ProjectTo{AbstractArray}(; element = ProjectTo(zero(T)), axes = axes(x))
-end
-(p::ProjectTo{AbstractArray{T}})(x::AbstractArray{T}) where {T <: TaylorScalar} = x
-accum_sum(xs::AbstractArray{T}; dims = :) where {T <: TaylorScalar} = sum(xs, dims = dims)
-
-TaylorNumeric{T <: TaylorScalar} = Union{T, AbstractArray{<:T}}
-
-@adjoint function broadcasted(::typeof(+), xs::TaylorNumeric...)
-    broadcast(+, xs...), ȳ -> (nothing, map(x -> unbroadcast(x, ȳ), xs)...)
-end
+(project::ProjectTo{T})(dx::TaylorScalar{T, N}) where {N, T <: Number} = primal(dx)
 
-struct TaylorOneElement{T, N, I, A} <: AbstractArray{T, N}
-    val::T
-    ind::I
-    axes::A
-    function TaylorOneElement(val::T, ind::I,
-            axes::A) where {T <: TaylorScalar, I <: NTuple{N, Int},
-            A <: NTuple{N, AbstractUnitRange}} where {N}
-        new{T, N, I, A}(val, ind, axes)
-    end
-end
+# opt-outs
 
-Base.size(A::TaylorOneElement) = map(length, A.axes)
-Base.axes(A::TaylorOneElement) = A.axes
-function Base.getindex(A::TaylorOneElement{T, N}, i::Vararg{Int, N}) where {T, N}
-    ifelse(i == A.ind, A.val, zero(T))
-end
-
-function ∇getindex(x::AbstractArray{T, N}, inds) where {T <: TaylorScalar, N}
-    dy -> begin
-        dx = TaylorOneElement(dy, inds, axes(x))
-        return (_project(x, dx), map(_ -> nothing, inds)...)
-    end
-end
+# Unary functions
 
-@generated function mul_adjoint(Ω::TaylorScalar{T, N}, x::TaylorScalar{T, N}) where {T, N}
-    return quote
-        vΩ, vx = value(Ω), value(x)
-        @inbounds TaylorScalar($([:(+($([:($(binomial(j - 1, i - 1)) * vΩ[$j] *
-                                           vx[$(j + 1 - i)]) for j in i:N]...)))
-                                  for i in 1:N]...))
-    end
+for f in (
+    exp, exp10, exp2, expm1,
+    sin, cos, tan, sec, csc, cot,
+    sinh, cosh, tanh, sech, csch, coth,
+    log, log10, log2, log1p,
+    asin, acos, atan, asec, acsc, acot,
+    asinh, acosh, atanh, asech, acsch, acoth,
+    sqrt, cbrt, inv
+)
+    @eval @opt_out frule(::typeof($f), x::TaylorScalar)
+    @eval @opt_out rrule(::typeof($f), x::TaylorScalar)
 end
 
-rrule(::typeof(*), x::TaylorScalar) = rrule(identity, x)
-
-function rrule(::typeof(*), x::TaylorScalar, y::TaylorScalar)
-    function times_pullback2(Ω̇)
-        ΔΩ = unthunk(Ω̇)
-        return (NoTangent(), ProjectTo(x)(mul_adjoint(ΔΩ, y)),
-            ProjectTo(y)(mul_adjoint(ΔΩ, x)))
-    end
-    return x * y, times_pullback2
-end
+# Binary functions
 
-function rrule(::typeof(*), x::TaylorScalar, y::TaylorScalar, z::TaylorScalar,
-        more::TaylorScalar...)
-    Ω2, back2 = rrule(*, x, y)
-    Ω3, back3 = rrule(*, Ω2, z)
-    Ω4, back4 = rrule(*, Ω3, more...)
-    function times_pullback4(Ω̇)
-        Δ4 = back4(unthunk(Ω̇))  # (0, ΔΩ3, Δmore...)
-        Δ3 = back3(Δ4[2])       # (0, ΔΩ2, Δz)
-        Δ2 = back2(Δ3[2])       # (0, Δx, Δy)
-        return (Δ2..., Δ3[3], Δ4[3:end]...)
+for f in (
+    *, /, ^
+)
+    for (tlhs, trhs) in (
+        (TaylorScalar, TaylorScalar),
+        (TaylorScalar, Number),
+        (Number, TaylorScalar)
+    )
+        @eval @opt_out frule(::typeof($f), x::$tlhs, y::$trhs)
+        @eval @opt_out rrule(::typeof($f), x::$tlhs, y::$trhs)
     end
-    return Ω4, times_pullback4
 end

test/runtests.jl

@@ -3,5 +3,5 @@ using Test
 
 include("primitive.jl")
 include("derivative.jl")
-# include("zygote.jl")
+include("zygote.jl")
 # include("lux.jl")

test/zygote.jl

@@ -1,38 +1,36 @@
-using Zygote, LinearAlgebra
+using LinearAlgebra
+import Zygote # use qualified import to avoid conflict with TaylorDiff
 
-@testset "Zygote for mixed derivative" begin
+@testset "Zygote-over-TaylorDiff on same variable" begin
+    # Scalar functions
     some_number = 0.7
     some_numbers = [0.3 0.4 0.1;]
-    for f in (exp, log, sqrt, sin, asin, sinh, asinh)
-        @test gradient(x -> derivative(f, x, 2), some_number)[1] ≈
+    for f in (exp, log, sqrt, sin, asin, sinh, asinh, x -> x^3)
+        @test Zygote.gradient(derivative, f, some_number, 2)[2] ≈
               derivative(f, some_number, 3)
-        derivative_result = vec(derivative.(f, some_numbers, 3))
-        @test Zygote.jacobian(x -> derivative.(f, x, 2), some_numbers)[1] ≈
-              diagm(derivative_result)
+        @test Zygote.jacobian(broadcast, derivative, f, some_numbers, 2)[3] ≈
+              diagm(vec(derivative.(f, some_numbers, 3)))
     end
 
-    some_matrix = [0.7 0.1; 0.4 0.2]
-    f = x -> sum(tanh.(x), dims = 1)
-    dfdx1(m, x) = derivative(m, x, [1.0, 0.0], 1)
-    dfdx2(m, x) = derivative(m, x, [0.0, 1.0], 1)
-    res(m, x) = dfdx1(m, x) .+ 2 * dfdx2(m, x)
-    grads = Zygote.gradient(some_matrix) do x
-        sum(res(f, x))
-    end
-    expected_grads = x -> -2 * sinh(x) / cosh(x)^3
-    @test grads[1] ≈ [1 0; 0 2] * expected_grads.(some_matrix)
-
-    @test gradient(x -> derivative(x -> x * x, x, 1),
-        5.0)[1] ≈ 2.0
-
+    # Vector functions
     g(x) = x[1] * x[1] + x[2] * x[2]
-    @test gradient(x -> derivative(g, x, [1.0, 0.0], 1),
-        [1.0, 2.0])[1] ≈ [2.0, 0.0]
+    @test Zygote.gradient(derivative, g, [1.0, 2.0], [1.0, 0.0], 1)[2] ≈ [2.0, 0.0]
+
+    # Matrix functions
+    some_matrix = [0.7 0.1; 0.4 0.2]
+    f(x) = sum(exp.(x), dims = 1)
+    dfdx1(x) = derivative(f, x, [1.0, 0.0], 1)
+    dfdx2(x) = derivative(f, x, [0.0, 1.0], 1)
+    res(x) = sum(dfdx1(x) .+ 2 * dfdx2(x))
+    grads = Zygote.gradient(res, some_matrix)
+    @test grads[1] ≈ [1 0; 0 2] * exp.(some_matrix)
 end
 
-@testset "Zygote for parameter optimization" begin
-    gradient(p -> derivative(x -> sum(exp.(x + p)), [1.0, 1.0], [1.0, 0.0], 1), [0.5, 0.7])
-    gradient(p -> derivative(x -> sum(exp.(p + x)), [1.0, 1.0], [1.0, 0.0], 1), [0.5, 0.7])
+@testset "Zygote-over-TaylorDiff on different variable" begin
+    Zygote.gradient(
+        p -> derivative(x -> sum(exp.(x + p)), [1.0, 1.0], [1.0, 0.0], 1), [0.5, 0.7])
+    Zygote.gradient(
+        p -> derivative(x -> sum(exp.(p + x)), [1.0, 1.0], [1.0, 0.0], 1), [0.5, 0.7])
     linear_model(x, p, b) = exp.(b + p * x + b)[1]
     some_x, some_v, some_p, some_b = [0.58, 0.36], [0.23, 0.11], [0.49 0.96], [0.88]
     loss_taylor(p) = derivative(x -> linear_model(x, p, some_b), some_x, some_v, 1)
@@ -41,5 +39,5 @@ end
         let f = x -> linear_model(x, p, some_b)
             (f(some_x + ε * some_v) - f(some_x - ε * some_v)) / 2ε
         end
-    @test gradient(loss_taylor, some_p)[1] ≈ gradient(loss_finite, some_p)[1]
+    @test Zygote.gradient(loss_taylor, some_p)[1] ≈ Zygote.gradient(loss_finite, some_p)[1]
 end