Fix gradient issues with kernelmatrix_diag and use ChainRulesCore (#208)

Co-authored-by: David Widmann <[email protected]>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

Author: 3 people

Project.toml

@@ -1,8 +1,9 @@
 name = "KernelFunctions"
 uuid = "ec8451be-7e33-11e9-00cf-bbf324bd1392"
-version = "0.8.25"
+version = "0.8.26"
 
 [deps]
+ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 Compat = "34da2185-b29b-5c13-b0c7-acf172513d20"
 Distances = "b4f34e82-e78d-54a5-968a-f98e89d6e8f7"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
@@ -17,8 +18,9 @@ Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 ZygoteRules = "700de1a5-db45-46bc-99cf-38207098b444"
 
 [compat]
+ChainRulesCore = "0.9"
 Compat = "3.7"
-Distances = "0.9.1, 0.10"
+Distances = "0.10"
 Functors = "0.1"
 Requires = "1.0.1"
 SpecialFunctions = "0.8, 0.9, 0.10, 1"

src/KernelFunctions.jl

@@ -55,12 +55,14 @@ export IndependentMOKernel, LatentFactorMOKernel
 export tensor, ⊗
 
 using Compat
+using ChainRulesCore: ChainRulesCore, Composite, Zero, One, DoesNotExist, NO_FIELDS
+using ChainRulesCore: @thunk, InplaceableThunk
 using Requires
 using Distances, LinearAlgebra
 using Functors
 using SpecialFunctions: loggamma, besselk, polygamma
-using ZygoteRules: @adjoint, pullback
-using StatsFuns: logtwo
+using ZygoteRules: ZygoteRules
+using StatsFuns: logtwo, twoπ
 using StatsBase
 using TensorCore
 
@@ -112,7 +114,8 @@ include(joinpath("mokernels", "moinput.jl"))
 include(joinpath("mokernels", "independent.jl"))
 include(joinpath("mokernels", "slfm.jl"))
 
-include("zygote_adjoints.jl")
+include("chainrules.jl")
+include("zygoterules.jl")
 
 include("test_utils.jl")
 

src/basekernels/fbm.jl

@@ -69,3 +69,14 @@ function kernelmatrix!(
     K .= _fbm.(_mod(x), _mod(y)', K, κ.h)
     return K
 end
+
+function kernelmatrix_diag(κ::FBMKernel, x::AbstractVector)
+    modx = _mod(x)
+    modxx = colwise(SqEuclidean(), x)
+    return _fbm.(modx, modx, modxx, κ.h)
+end
+
+function kernelmatrix_diag(κ::FBMKernel, x::AbstractVector, y::AbstractVector)
+    modxy = colwise(SqEuclidean(), x, y)
+    return _fbm.(_mod(x), _mod(y), modxy, κ.h)
+end

src/basekernels/gabor.jl

@@ -72,3 +72,7 @@ function kernelmatrix(κ::GaborKernel, x::AbstractVector, y::AbstractVector)
 end
 
 kernelmatrix_diag(κ::GaborKernel, x::AbstractVector) = kernelmatrix_diag(κ.kernel, x)
+
+function kernelmatrix_diag(κ::GaborKernel, x::AbstractVector, y::AbstractVector)
+    return kernelmatrix_diag(κ.kernel, x, y)
+end

src/basekernels/nn.jl

@@ -51,6 +51,19 @@ function kernelmatrix(::NeuralNetworkKernel, x::ColVecs)
     return asin.(XX ./ sqrt.(X_2_1' * X_2_1))
 end
 
+function kernelmatrix_diag(::NeuralNetworkKernel, x::ColVecs)
+    x_2 = vec(sum(x.X .* x.X; dims=1))
+    return asin.(x_2 ./ (x_2 .+ 1))
+end
+
+function kernelmatrix_diag(::NeuralNetworkKernel, x::ColVecs, y::ColVecs)
+    validate_inputs(x, y)
+    x_2 = vec(sum(x.X .* x.X; dims=1) .+ 1)
+    y_2 = vec(sum(y.X .* y.X; dims=1) .+ 1)
+    xy = vec(sum(x.X' .* y.X'; dims=2))
+    return asin.(xy ./ sqrt.(x_2 .* y_2))
+end
+
 function kernelmatrix(::NeuralNetworkKernel, x::RowVecs, y::RowVecs)
     validate_inputs(x, y)
     X_2 = sum(x.X .* x.X; dims=2)
@@ -65,4 +78,17 @@ function kernelmatrix(::NeuralNetworkKernel, x::RowVecs)
     return asin.(XX ./ sqrt.(X_2_1 * X_2_1'))
 end
 
+function kernelmatrix_diag(::NeuralNetworkKernel, x::RowVecs)
+    x_2 = vec(sum(x.X .* x.X; dims=2))
+    return asin.(x_2 ./ (x_2 .+ 1))
+end
+
+function kernelmatrix_diag(::NeuralNetworkKernel, x::RowVecs, y::RowVecs)
+    validate_inputs(x, y)
+    x_2 = vec(sum(x.X .* x.X; dims=2) .+ 1)
+    y_2 = vec(sum(y.X .* y.X; dims=2) .+ 1)
+    xy = vec(sum(x.X .* y.X; dims=2))
+    return asin.(xy ./ sqrt.(x_2 .* y_2))
+end
+
 Base.show(io::IO, ::NeuralNetworkKernel) = print(io, "Neural Network Kernel")

src/chainrules.jl

@@ -0,0 +1,167 @@
+## Forward Rules
+
+# Note that this is type piracy as the derivative should be NaN for x == y.
+function ChainRulesCore.frule(
+    (_, Δx, Δy), d::Distances.Euclidean, x::AbstractVector, y::AbstractVector
+)
+    Δ = x - y
+    D = sqrt(sum(abs2, Δ))
+    if !iszero(D)
+        Δ ./= D
+    end
+    return D, dot(Δ, Δx) - dot(Δ, Δy)
+end
+
+## Reverse Rules Delta
+
+function ChainRulesCore.rrule(dist::Delta, x::AbstractVector, y::AbstractVector)
+    d = dist(x, y)
+    function evaluate_pullback(::Any)
+        return NO_FIELDS, Zero(), Zero()
+    end
+    return d, evaluate_pullback
+end
+
+function ChainRulesCore.rrule(
+    ::typeof(Distances.pairwise), d::Delta, X::AbstractMatrix, Y::AbstractMatrix; dims=2
+)
+    P = Distances.pairwise(d, X, Y; dims=dims)
+    function pairwise_pullback(::AbstractMatrix)
+        return NO_FIELDS, NO_FIELDS, Zero(), Zero()
+    end
+    return P, pairwise_pullback
+end
+
+function ChainRulesCore.rrule(
+    ::typeof(Distances.pairwise), d::Delta, X::AbstractMatrix; dims=2
+)
+    P = Distances.pairwise(d, X; dims=dims)
+    function pairwise_pullback(::AbstractMatrix)
+        return NO_FIELDS, NO_FIELDS, Zero()
+    end
+    return P, pairwise_pullback
+end
+
+function ChainRulesCore.rrule(
+    ::typeof(Distances.colwise), d::Delta, X::AbstractMatrix, Y::AbstractMatrix
+)
+    C = Distances.colwise(d, X, Y)
+    function colwise_pullback(::AbstractVector)
+        return NO_FIELDS, NO_FIELDS, Zero(), Zero()
+    end
+    return C, colwise_pullback
+end
+
+## Reverse Rules DotProduct
+
+function ChainRulesCore.rrule(dist::DotProduct, x::AbstractVector, y::AbstractVector)
+    d = dist(x, y)
+    function evaluate_pullback(Δ::Any)
+        return NO_FIELDS, Δ .* y, Δ .* x
+    end
+    return d, evaluate_pullback
+end
+
+function ChainRulesCore.rrule(
+    ::typeof(Distances.pairwise),
+    d::DotProduct,
+    X::AbstractMatrix,
+    Y::AbstractMatrix;
+    dims=2,
+)
+    P = Distances.pairwise(d, X, Y; dims=dims)
+    function pairwise_pullback_cols(Δ::AbstractMatrix)
+        if dims == 1
+            return NO_FIELDS, NO_FIELDS, Δ * Y, Δ' * X
+        else
+            return NO_FIELDS, NO_FIELDS, Y * Δ', X * Δ
+        end
+    end
+    return P, pairwise_pullback_cols
+end
+
+function ChainRulesCore.rrule(
+    ::typeof(Distances.pairwise), d::DotProduct, X::AbstractMatrix; dims=2
+)
+    P = Distances.pairwise(d, X; dims=dims)
+    function pairwise_pullback_cols(Δ::AbstractMatrix)
+        if dims == 1
+            return NO_FIELDS, NO_FIELDS, 2 * Δ * X
+        else
+            return NO_FIELDS, NO_FIELDS, 2 * X * Δ
+        end
+    end
+    return P, pairwise_pullback_cols
+end
+
+function ChainRulesCore.rrule(
+    ::typeof(Distances.colwise), d::DotProduct, X::AbstractMatrix, Y::AbstractMatrix
+)
+    C = Distances.colwise(d, X, Y)
+    function colwise_pullback(Δ::AbstractVector)
+        return NO_FIELDS, NO_FIELDS, Δ' .* Y, Δ' .* X
+    end
+    return C, colwise_pullback
+end
+
+## Reverse Rules Sinus
+
+function ChainRulesCore.rrule(s::Sinus, x::AbstractVector, y::AbstractVector)
+    d = x - y
+    sind = sinpi.(d)
+    abs2_sind_r = abs2.(sind) ./ s.r
+    val = sum(abs2_sind_r)
+    gradx = twoπ .* cospi.(d) .* sind ./ (s.r .^ 2)
+    function evaluate_pullback(Δ::Any)
+        return (r=-2Δ .* abs2_sind_r,), Δ * gradx, -Δ * gradx
+    end
+    return val, evaluate_pullback
+end
+
+## Reverse Rulse SqMahalanobis
+
+function ChainRulesCore.rrule(
+    dist::Distances.SqMahalanobis, a::AbstractVector, b::AbstractVector
+)
+    d = dist(a, b)
+    function SqMahalanobis_pullback(Δ::Real)
+        a_b = a - b
+        ∂qmat = InplaceableThunk(
+            @thunk((a_b * a_b') * Δ), X̄ -> mul!(X̄, a_b, a_b', true, Δ)
+        )
+        ∂a = InplaceableThunk(
+            @thunk((2 * Δ) * dist.qmat * a_b), X̄ -> mul!(X̄, dist.qmat, a_b, true, 2 * Δ)
+        )
+        ∂b = InplaceableThunk(
+            @thunk((-2 * Δ) * dist.qmat * a_b), X̄ -> mul!(X̄, dist.qmat, a_b, true, -2 * Δ)
+        )
+        return Composite{typeof(dist)}(; qmat=∂qmat), ∂a, ∂b
+    end
+    return d, SqMahalanobis_pullback
+end
+
+## Reverse Rules for matrix wrappers
+
+function ChainRulesCore.rrule(::Type{<:ColVecs}, X::AbstractMatrix)
+    ColVecs_pullback(Δ::Composite) = (NO_FIELDS, Δ.X)
+    function ColVecs_pullback(::AbstractVector{<:AbstractVector{<:Real}})
+        return error(
+            "Pullback on AbstractVector{<:AbstractVector}.\n" *
+            "This might happen if you try to use gradients on the generic `kernelmatrix` or `kernelmatrix_diag`.\n" *
+            "To solve this issue overload `kernelmatrix(_diag)` for your kernel for `ColVecs`",
+        )
+    end
+    return ColVecs(X), ColVecs_pullback
+end
+
+function ChainRulesCore.rrule(::Type{<:RowVecs}, X::AbstractMatrix)
+    RowVecs_pullback(Δ::Composite) = (NO_FIELDS, Δ.X)
+    function RowVecs_pullback(::AbstractVector{<:AbstractVector{<:Real}})
+        return error(
+            "Pullback on AbstractVector{<:AbstractVector}.\n" *
+            "This might happen if you try to use gradients on the generic `kernelmatrix` or `kernelmatrix_diag`.\n" *
+            "To solve this issue overload `kernelmatrix(_diag)` for your kernel for `RowVecs`",
+        )
+    end
+    return RowVecs(X), RowVecs_pullback
+end

src/distances/delta.jl

@@ -1,4 +1,5 @@
-struct Delta <: Distances.PreMetric end
+# Delta is not following the PreMetric rules since d(x, x) == 1
+struct Delta <: Distances.UnionPreMetric end
 
 @inline function Distances._evaluate(::Delta, a::AbstractVector, b::AbstractVector)
     @boundscheck if length(a) != length(b)
@@ -12,7 +13,7 @@ struct Delta <: Distances.PreMetric end
     return a == b
 end
 
-Distances.result_type(::Delta, Ta::Type, Tb::Type) = promote_type(Ta, Tb)
+Distances.result_type(::Delta, Ta::Type, Tb::Type) = Bool
 
 @inline (dist::Delta)(a::AbstractArray, b::AbstractArray) = Distances._evaluate(dist, a, b)
 @inline (dist::Delta)(a::Number, b::Number) = a == b

src/distances/dotproduct.jl

@@ -1,5 +1,5 @@
-struct DotProduct <: Distances.PreMetric end
-# struct DotProduct <: Distances.UnionSemiMetric end
+## DotProduct is not following the PreMetric rules since d(x, x) != 0 and d(x, y) >= 0 for all x, y
+struct DotProduct <: Distances.UnionPreMetric end
 
 @inline function Distances._evaluate(::DotProduct, a::AbstractVector, b::AbstractVector)
     @boundscheck if length(a) != length(b)

src/distances/pairwise.jl

@@ -29,3 +29,34 @@ function pairwise!(
 )
     return Distances.pairwise!(out, d, reshape(x, :, 1), reshape(y, :, 1); dims=1)
 end
+
+# Also defines the colwise method for abstractvectors
+
+function colwise(d::PreMetric, x::AbstractVector)
+    return zeros(Distances.result_type(d, x, x), length(x)) # Valid since d(x,x) == 0 by definition
+end
+
+## The following is a hack for DotProduct and Delta to still work
+function colwise(d::Distances.UnionPreMetric, x::ColVecs)
+    return Distances.colwise(d, x.X, x.X)
+end
+
+function colwise(d::Distances.UnionPreMetric, x::RowVecs)
+    return Distances.colwise(d, x.X', x.X')
+end
+
+function colwise(d::Distances.UnionPreMetric, x::AbstractVector)
+    return map(d, x, x)
+end
+
+function colwise(d::PreMetric, x::ColVecs, y::ColVecs)
+    return Distances.colwise(d, x.X, y.X)
+end
+
+function colwise(d::PreMetric, x::RowVecs, y::RowVecs)
+    return Distances.colwise(d, x.X', y.X')
+end
+
+function colwise(d::PreMetric, x::AbstractVector, y::AbstractVector)
+    return map(d, x, y)
+end

src/distances/sinus.jl

@@ -1,5 +1,4 @@
-struct Sinus{T} <: Distances.SemiMetric
-    # struct Sinus{T} <: Distances.UnionSemiMetric
+struct Sinus{T} <: Distances.UnionSemiMetric
     r::Vector{T}
 end