Do not use internal PDMats.chol_lower (#133)

* Test optimization of variational covariance * Use ParameterHandling This is required for the new test * Add ParameterHandling * Include new test * Clean up test script * Limit PDMats to version 0.11.7 * Use explicit compat interval Co-authored-by: David Widmann <[email protected]> * Add/modify compate entries for tests * Do not use PDMats.chol_lower * Update Project.toml * Fix format Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Remove additional test Co-authored-by: Simone Carlo Surace <[email protected]> Co-authored-by: Simone Surace <[email protected]> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
JuliaGaussianProcesses · May 9, 2022 · a9c3ecd · a9c3ecd · devmotion · May 9, 2022
1 parent dec9785
commit a9c3ecd
Show file tree

Hide file tree

Showing 3 changed files with 10 additions and 8 deletions.
diff --git a/Project.toml b/Project.toml
@@ -1,7 +1,7 @@
 name = "ApproximateGPs"
 uuid = "298c2ebc-0411-48ad-af38-99e88101b606"
 authors = ["JuliaGaussianProcesses Team"]
-version = "0.4.1"
+version = "0.4.2"
 
 [deps]
 AbstractGPs = "99985d1d-32ba-4be9-9821-2ec096f28918"

diff --git a/src/SparseVariationalApproximationModule.jl b/src/SparseVariationalApproximationModule.jl
@@ -9,7 +9,7 @@ using LinearAlgebra
 using Statistics
 using StatsBase
 using FillArrays: Fill
-using PDMats: chol_lower, ScalMat
+using PDMats: ScalMat
 
 using AbstractGPs: AbstractGPs
 using AbstractGPs:
@@ -23,7 +23,7 @@ using AbstractGPs:
     At_A,
     diag_At_A
 using GPLikelihoods: GaussianLikelihood, DefaultExpectationMethod, expected_loglikelihood
-using ..ApproximateGPs: _chol_cov, _cov
+using ..ApproximateGPs: _chol_lower, _chol_cov, _cov
 
 @doc raw"""
     Centered()
@@ -129,7 +129,7 @@ function AbstractGPs.posterior(sva::SparseVariationalApproximation{Centered})
     q, fz = sva.q, sva.fz
     m, S = mean(q), _chol_cov(q)
     Kuu = _chol_cov(fz)
-    B = chol_lower(Kuu) \ chol_lower(S)
+    B = _chol_lower(Kuu) \ _chol_lower(S)
     α = Kuu \ (m - mean(fz))
     data = (Kuu=Kuu, B=B, α=α)
     return ApproxPosteriorGP(sva, fz.f, data)
@@ -179,9 +179,9 @@ function AbstractGPs.posterior(sva::SparseVariationalApproximation{NonCentered})
     q, fz = sva.q, sva.fz
     m = mean(q)
     Kuu = _chol_cov(fz)
-    α = chol_lower(Kuu)' \ m
+    α = _chol_lower(Kuu)' \ m
     Sv = _chol_cov(q)
-    B = chol_lower(Sv)
+    B = _chol_lower(Sv)
     data = (Kuu=Kuu, B=B, α=α)
     return ApproxPosteriorGP(sva, fz.f, data)
 end
@@ -214,7 +214,7 @@ end
 # A = Lk⁻¹ Ku* is the projection matrix used in computing the predictive variance of the SparseVariationalApproximation posterior.
 function _A_and_Kuf(f, x)
     Kuf = cov(f.prior, inducing_points(f), x)
-    A = chol_lower(f.data.Kuu) \ Kuf
+    A = _chol_lower(f.data.Kuu) \ Kuf
     return A, Kuf
 end
 
@@ -366,7 +366,7 @@ function _prior_kl(sva::SparseVariationalApproximation{NonCentered})
     C_ε = _cov(sva.q)
 
     # trace_term = tr(C_ε)  # does not work due to PDMat / Zygote issues
-    L = chol_lower(_chol_cov(sva.q))
+    L = _chol_lower(_chol_cov(sva.q))
     trace_term = sum(L .^ 2)  # TODO remove AD workaround
 
     return (trace_term + m_ε'm_ε - length(m_ε) - logdet(C_ε)) / 2

diff --git a/src/utils.jl b/src/utils.jl
@@ -12,6 +12,8 @@ if VERSION < v"1.7"
     end
 end
 
+_chol_lower(a::Cholesky) = a.uplo === 'L' ? a.L : a.U'
+
 _chol_cov(q::AbstractMvNormal) = cholesky(Symmetric(cov(q)))
 _chol_cov(q::MvNormal) = cholesky(q.Σ)
-Original file line number
+Diff line change
@@ Expand Up / @@ -12,6 +12,8 @@ if VERSION < v"1.7" @@
         end
     end
+    _chol_lower(a::Cholesky) = a.uplo === 'L' ? a.L : a.U'
     _chol_cov(q::AbstractMvNormal) = cholesky(Symmetric(cov(q)))
     _chol_cov(q::MvNormal) = cholesky(q.Σ)
@@ Expand Down @@