remove @adjoint function cholesky

.github/workflows/ci.yml

@@ -17,7 +17,7 @@ jobs:
       fail-fast: false
       matrix:
         version:
-          - '1.3' # Replace this with the minimum Julia version that your package supports.
+          - '1.6' # Replace this with the minimum Julia version that your package supports.
           - '1'   # automatically expands to the latest stable 1.x release of Julia
           - 'nightly'
         os:

Project.toml

@@ -25,7 +25,7 @@ ZygoteRules = "700de1a5-db45-46bc-99cf-38207098b444"
 
 [compat]
 AbstractFFTs = "0.5, 1.0"
-ChainRules = "1.5"
+ChainRules = "1.35.3"
 ChainRulesCore = "1.9"
 ChainRulesTestUtils = "1"
 DiffRules = "1.4"
@@ -38,7 +38,7 @@ NaNMath = "0.3, 1"
 Requires = "1.1"
 SpecialFunctions = "1.6, 2"
 ZygoteRules = "0.2.1"
-julia = "1.3"
+julia = "1.6"
 
 [extras]
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"

src/lib/array.jl

@@ -564,35 +564,6 @@ end
 @adjoint Matrix(A::LinearAlgebra.HermOrSym{T,S}) where {T,S} = Matrix(A),
   Δ -> (convert(S, Δ),)
 
-@adjoint function cholesky(Σ::Real)
-  C = cholesky(Σ)
-  return C, Δ::NamedTuple->(Δ.factors[1, 1] / (2 * C.U[1, 1]),)
-end
-
-@adjoint function cholesky(Σ::Diagonal; check = true)
-  C = cholesky(Σ, check = check)
-  return C, Δ::NamedTuple -> begin
-    issuccess(C) || throw(PosDefException(C.info))
-    return Diagonal(diag(Δ.factors) .* inv.(2 .* C.factors.diag)), nothing
-  end
-end
-
-# Implementation due to Seeger, Matthias, et al. "Auto-differentiating linear algebra."
-@adjoint function cholesky(Σ::Union{StridedMatrix, Symmetric{<:Real, <:StridedMatrix}}; check = true)
-  C = cholesky(Σ, check = check)
-  return C, function(Δ::NamedTuple)
-    issuccess(C) || throw(PosDefException(C.info))
-    U, Ū = C.U, Δ.factors
-    Σ̄ = similar(U.data)
-    Σ̄ = mul!(Σ̄, Ū, U')
-    Σ̄ = copytri!(Σ̄, 'U')
-    Σ̄ = ldiv!(U, Σ̄)
-    Σ̄ = BLAS.trsm!('R', 'U', 'T', 'N', one(eltype(Σ)), U.data, Σ̄)
-    Σ̄[diagind(Σ̄)] ./= 2
-    return (UpperTriangular(Σ̄),)
-  end
-end
-
 @adjoint function lyap(A::AbstractMatrix, C::AbstractMatrix)
   X = lyap(A, C)
   return X, function (X̄)

test/gradcheck.jl

@@ -654,7 +654,6 @@ end
     g(X) = cholesky(X * X' + I)
     @test Zygote.pullback(g, X)[2]((factors=LowerTriangular(X),))[1] ≈
       Zygote.pullback(g, X)[2]((factors=Matrix(LowerTriangular(X)),))[1]
-    @test_throws PosDefException Zygote.pullback(X -> cholesky(X, check = false), X)[2]((factors=X,))
 
     # https://github.com/FluxML/Zygote.jl/issues/932
     @test gradcheck(rand(5, 5), rand(5)) do A, x
@@ -820,6 +819,32 @@ end
     @test back′(C̄)[1] isa Diagonal
     @test diag(back′(C̄)[1]) ≈ diag(back(C̄)[1])
   end
+  @testset "cholesky - Hermitian{Complex}" begin
+    rng, N = MersenneTwister(123456), 3
+    A = randn(rng, Complex{Float64}, N, N)
+    H = Hermitian(A * A' + I)
+    Hmat = Matrix(H)
+    y, back = Zygote.pullback(cholesky, Hmat)
+    y′, back′ = Zygote.pullback(cholesky, H)
+    C̄ = (factors=randn(rng, N, N),)
+    @test only(back′(C̄)) isa Hermitian
+    # gradtest does not support complex gradients, even though the pullback exists
+    d = only(back(C̄))
+    d′ = only(back′(C̄))
+    @test (d + d')/2 ≈ d′
+  end
+  @testset "cholesky - Hermitian{Real}" begin
+    rng, N = MersenneTwister(123456), 3
+    A = randn(rng, N, N)
+    H = Hermitian(A * A' + I)
+    Hmat = Matrix(H)
+    y, back = Zygote.pullback(cholesky, Hmat)
+    y′, back′ = Zygote.pullback(cholesky, H)
+    C̄ = (factors=randn(rng, N, N),)
+    @test back′(C̄)[1] isa Hermitian
+    @test gradtest(B->cholesky(Hermitian(B)).U, Hmat)
+    @test gradtest(B->logdet(cholesky(Hermitian(B))), Hmat)
+  end
 end
 
 @testset "lyap" begin