used new style of transform for Kernelfunctions (#28)

* used new style of transform for Kernelfunctions

* update version number KernelFunctions

Co-authored-by: Johannes Giersdorf <johannes.giersdorf@gmail.com>

Project.toml

@@ -14,6 +14,6 @@ Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 [compat]
 AbstractMCMC = "3"
 Distributions = "0.24, 0.25"
-KernelFunctions = "0.8, 0.9"
+KernelFunctions = "0.10"
 PDMats = "0.10, 0.11"
 julia = "1.3"

src/bayesquads/bayesquad.jl

@@ -38,11 +38,11 @@ function BayesQuad(k::ScaledKernel; l=1.0, σ=nothing)
 end
 
 function kernel(b::BayesQuad)
-    return b.σ * transform(b.kernel, inv.(b.l))
+    return b.σ * (b.kernel ∘ ScaleTransform(inv.(b.l)))
 end
 
 function kernel(b::BayesQuad{<:Kernel,<:LowerTriangular})
-    return b.σ * transform(b.kernel, LinearTransform(inv(b.l)))
+    return b.σ * (b.kernel ∘ LinearTransform(inv(b.l)))
 end
 
 function quadrature(

test/bayesquads/bayesquad.jl

@@ -1,3 +1,9 @@
 @testset "bayesquad" begin
-
+    @testset "lengthscale and variance" begin
+        σ²_k = 1.5; l_k = 1.5
+        k = σ²_k * SqExponentialKernel() ∘ ScaleTransform(1. /  l_k)
+        bq = BayesQuad(k)
+        @test bq.σ ≈ σ²_k
+        @test bq.l ≈ l_k
+    end
 end

test/interface.jl

@@ -6,7 +6,7 @@
     logintegrand = x->logpdf(likelihood, x)
     m = BayesModel(prior, logintegrand)
     s = PriorSampling()
-    bquad = BayesQuad(transform(SEKernel(), 0.1))
+    bquad = BayesQuad(SEKernel() ∘ ScaleTransform(0.1))
     pZ, x  = bquad(m, s; nsamples=100)
     trueZ = exp(-0.5 * (logdet(cov(prior) + cov(likelihood)) + D * log(2π)))
     @test mean(pZ) ≈ trueZ atol=1e-3