Remove Symmetric(::Diagonal) type piracy

Project.toml

@@ -1,7 +1,7 @@
 name = "AbstractGPs"
 uuid = "99985d1d-32ba-4be9-9821-2ec096f28918"
 authors = ["JuliaGaussianProcesses Team"]
-version = "0.2.19"
+version = "0.2.20"
 
 [deps]
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"

src/abstract_gp/finite_gp.jl

@@ -183,7 +183,7 @@ true
 """
 function Random.rand(rng::AbstractRNG, f::FiniteGP, N::Int)
     m, C_mat = mean_and_cov(f)
-    C = cholesky(Symmetric(C_mat))
+    C = cholesky(_symmetric(C_mat))
     return m .+ C.U' * randn(rng, promote_type(eltype(m), eltype(C)), length(m), N)
 end
 Random.rand(f::FiniteGP, N::Int) = rand(Random.GLOBAL_RNG, f, N)
@@ -223,7 +223,7 @@ Distributions.loglikelihood(f::FiniteGP, Y::AbstractMatrix{<:Real}) = sum(logpdf
 
 function Distributions.logpdf(f::FiniteGP, Y::AbstractMatrix{<:Real})
     m, C_mat = mean_and_cov(f)
-    C = cholesky(Symmetric(C_mat))
+    C = cholesky(_symmetric(C_mat))
     T = promote_type(eltype(m), eltype(C), eltype(Y))
     return -((size(Y, 1) * T(log(2π)) + logdet(C)) .+ diag_Xt_invA_X(C, Y .- m)) ./ 2
 end
@@ -285,19 +285,12 @@ function dtc(f::FiniteGP, y::AbstractVector{<:Real}, u::FiniteGP)
     return first(_compute_intermediates(f, y, u))
 end
 
-# Small bit of indirection to work around a cholesky-related bug whereby the interaction
-# between `FillArrays` and `Diagonal` and `Cholesky` causes problems.
-_cholesky(X) = cholesky(X)
-function _cholesky(X::Diagonal{<:Real, <:FillArrays.AbstractFill})
-    return cholesky(Diagonal(collect(diag(X))))
-end
-
 # Factor out computations common to the `elbo` and `dtc`.
 function _compute_intermediates(f::FiniteGP, y::AbstractVector{<:Real}, u::FiniteGP)
     consistency_check(f, y, u)
     chol_Σy = _cholesky(f.Σy)
 
-    A = cholesky(Symmetric(cov(u))).U' \ (chol_Σy.U' \ cov(f, u))'
+    A = cholesky(_symmetric(cov(u))).U' \ (chol_Σy.U' \ cov(f, u))'
     Λ_ε = cholesky(Symmetric(A * A' + I))
     δ = chol_Σy.U' \ (y - mean(f))
 

src/posterior_gp/approx_posterior_gp.jl

@@ -7,12 +7,6 @@ struct ApproxPosteriorGP{Tapprox, Tprior, Tdata} <: AbstractGP
     data::Tdata
 end
 
-# If a matrix is `Diagonal`, we generally don't need to wrap it in a `Symmetric`, because
-# it's already symmetric. This is used in a couple of places to avoid precisely this and
-# having to add specialised methods of e.g. `_cholesky` for complicated wrapped types.
-_symmetric(X) = Symmetric(X)
-_symmetric(X::Diagonal) = X
-
 """
     approx_posterior(::VFE, fx::FiniteGP, y::AbstractVector{<:Real}, u::FiniteGP)
 
@@ -25,14 +19,14 @@ Intelligence and Statistics. 2009.
 """
 function approx_posterior(::VFE, fx::FiniteGP, y::AbstractVector{<:Real}, u::FiniteGP)
     U_y = _cholesky(_symmetric(fx.Σy)).U
-    U = cholesky(Symmetric(cov(u))).U
+    U = cholesky(_symmetric(cov(u))).U
 
     B_εf = U' \ (U_y' \ cov(fx, u))'
 
     b_y = U_y' \ (y - mean(fx))
 
     D = B_εf * B_εf' + I
-    Λ_ε = cholesky(Symmetric(D))
+    Λ_ε = cholesky(_symmetric(D))
 
     m_ε = Λ_ε \ (B_εf * b_y)
 
@@ -117,7 +111,7 @@ function update_approx_posterior(
 )
     U11 = f_post_approx.data.U
     C12 = cov(u.f, f_post_approx.data.z, u.x)
-    C22 = Symmetric(cov(u))
+    C22 = _symmetric(cov(u))
     U = update_chol(Cholesky(U11,'U', 0), C12, C22).U
     U22 = U[end-length(u)+1:end, end-length(u)+1:end]
     U12 = U[1:length(f_post_approx.data.z), end-length(u)+1:end]
@@ -154,8 +148,6 @@ function update_approx_posterior(
     return ApproxPosteriorGP(VFE(), f_post_approx.prior, cache)
 end
 
-LinearAlgebra.Symmetric(X::Diagonal) = X
-
 # AbstractGP interface implementation.
 
 function Statistics.mean(f::ApproxPosteriorGP{VFE}, x::AbstractVector)

src/posterior_gp/posterior_gp.jl

@@ -20,7 +20,7 @@ where `m` and `k` are the mean function and kernel of `fx.f` respectively.
 """
 function posterior(fx::FiniteGP, y::AbstractVector{<:Real})
     m, C_mat = mean_and_cov(fx)
-    C = cholesky(Symmetric(C_mat))
+    C = cholesky(_symmetric(C_mat))
     δ = y - m
     α = C \ δ
     return PosteriorGP(fx.f, (α=α, C=C, x=fx.x, δ=δ))

src/util/common_covmat_ops.jl

@@ -1,3 +1,15 @@
+# If a matrix is `Diagonal`, we generally don't need to wrap it in a `Symmetric`, because
+# it's already symmetric. This is used in a couple of places to avoid precisely this and
+# having to add specialised methods of e.g. `_cholesky` for complicated wrapped types.
+_symmetric(X) = Symmetric(X)
+_symmetric(X::Diagonal) = X
+
+# Small bit of indirection to work around a cholesky-related bug whereby the interaction
+# between `FillArrays` and `Diagonal` and `Cholesky` causes problems.
+_cholesky(X) = cholesky(X)
+function _cholesky(X::Diagonal{<:Real, <:FillArrays.AbstractFill})
+    return cholesky(Diagonal(collect(diag(X))))
+end
 
 """
      update_chol(chol::Cholesky, C12::AbstractMatrix, C22::AbstractMatrix)

test/abstract_gp/finite_gp.jl

@@ -169,12 +169,6 @@ end
         #     atol=1e-8, rtol=1e-8,
         # )
 
-        # Supporting utility function.
-        @testset "_cholesky" begin
-            D = Diagonal(Fill(5.0, 10))
-            @test AbstractGPs._cholesky(D).U ≈ AbstractGPs._cholesky(collect(D)).U
-        end
-
         # Ensure that the elbo is close to the logpdf when appropriate.
         @test elbo(y, ŷ, fx) isa Real
         @test elbo(y, ŷ, fx) ≈ logpdf(y, ŷ)

test/posterior_gp/approx_posterior_gp.jl

@@ -1,25 +1,9 @@
 @testset "approx_posterior_gp" begin
-
-    @testset "_symmetric" begin
-        @testset "Matrix" begin
-            X = randn(5, 5)
-            @test AbstractGPs._symmetric(X) isa Symmetric
-            @test collect(AbstractGPs._symmetric(X)) == collect(Symmetric(X))
-        end
-        @testset "Diagonal" begin
-            X = Diagonal(randn(5))
-            @test AbstractGPs._symmetric(X) isa Diagonal
-            @test collect(AbstractGPs._symmetric(X)) == collect(Symmetric(X))
-        end
-    end
-
     rng = MersenneTwister(123456)
     N_cond = 3
     N_a = 5
     N_b = 6
 
-    @test Symmetric(Diagonal(randn(rng, 5))) isa Diagonal
-
     # Specify prior.
     f = GP(sin, Matern32Kernel())
 

test/util/common_covmat_ops.jl

@@ -1,4 +1,23 @@
 @testset "common_covmat_ops" begin
+    # Supporting utility functions.
+    @testset "_cholesky" begin
+        D = Diagonal(Fill(5.0, 10))
+        @test AbstractGPs._cholesky(D).U ≈ AbstractGPs._cholesky(collect(D)).U
+    end
+
+    @testset "_symmetric" begin
+        @testset "Matrix" begin
+            X = randn(5, 5)
+            @test AbstractGPs._symmetric(X) isa Symmetric
+            @test collect(AbstractGPs._symmetric(X)) == collect(Symmetric(X))
+        end
+        @testset "Diagonal" begin
+            X = Diagonal(randn(5))
+            @test AbstractGPs._symmetric(X) isa Diagonal
+            @test collect(AbstractGPs._symmetric(X)) == collect(Symmetric(X))
+        end
+    end
+
     @testset "update_chol" begin
         X = rand(5)
         y = rand(5)