create ApproximateGPs.TestUtils

Author: st--

src/ApproximateGPs.jl

@@ -23,4 +23,6 @@ include("LaplaceApproximationModule.jl")
 
 include("deprecations.jl")
 
+include("TestUtils.jl")
+
 end

src/TestUtils.jl

@@ -0,0 +1,74 @@
+module TestUtils
+
+using LinearAlgebra
+using Random
+using Test
+
+using Distributions
+using LogExpFunctions: logistic
+
+using AbstractGPs
+using ApproximateGPs
+
+function generate_data()
+    X = range(0, 23.5; length=48)
+    # The random number generator changed in 1.6->1.7. The following vector was generated in Julia 1.6.
+    # The generating code below is only kept for illustrative purposes.
+    #! format: off
+    Y = [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0]
+    #! format: on
+    # Random.seed!(1)
+    # fs = @. 3 * sin(10 + 0.6X) + sin(0.1X) - 1
+    # # invlink = normcdf
+    # invlink = logistic
+    # ps = invlink.(fs)
+    # Y = @. rand(Bernoulli(ps))
+    return X, Y
+end
+
+dist_y_given_f(f) = Bernoulli(logistic(f))
+
+function build_latent_gp(theta)
+    variance = softplus(theta[1])
+    lengthscale = softplus(theta[2])
+    kernel = variance * with_lengthscale(SqExponentialKernel(), lengthscale)
+    return LatentGP(GP(kernel), dist_y_given_f, 1e-8)
+end
+
+function test_approximation_predictions(approx)
+    rng = MersenneTwister(123456)
+    N_cond = 5
+    N_a = 6
+    N_b = 7
+
+    # Specify prior.
+    f = GP(Matern32Kernel())
+    # Sample from prior.
+    x = collect(range(-1.0, 1.0; length=N_cond))
+    noise_scale = 0.1
+    fx = f(x, noise_scale^2)
+    y = rand(rng, fx)
+
+    jitter = 0.0  # not needed in Gaussian case
+    lf = LatentGP(f, f -> Normal(f, noise_scale), jitter)
+    f_approx_post = posterior(approx, lf(x), y)
+
+    @testset "AbstractGPs API" begin
+        a = collect(range(-1.2, 1.2; length=N_a))
+        b = randn(rng, N_b)
+        AbstractGPs.TestUtils.test_internal_abstractgps_interface(rng, f_approx_post, a, b)
+    end
+
+    @testset "exact GPR equivalence for Gaussian likelihood" begin
+        f_exact_post = posterior(f(x, noise_scale^2), y)
+        xt = vcat(x, randn(rng, 3))  # test at training and new points
+
+        m_approx, c_approx = mean_and_cov(f_approx_post(xt))
+        m_exact, c_exact = mean_and_cov(f_exact_post(xt))
+
+        @test m_approx ≈ m_exact
+        @test c_approx ≈ c_exact
+    end
+end
+
+end

test/LaplaceApproximationModule.jl

@@ -1,28 +1,7 @@
 @testset "laplace" begin
-    function generate_data()
-        X = range(0, 23.5; length=48)
-        # The random number generator changed in 1.6->1.7. The following vector was generated in Julia 1.6.
-        # The generating code below is only kept for illustrative purposes.
-        #! format: off
-        Y = [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0]
-        #! format: on
-        # Random.seed!(1)
-        # fs = @. 3 * sin(10 + 0.6X) + sin(0.1X) - 1
-        # # invlink = normcdf
-        # invlink = logistic
-        # ps = invlink.(fs)
-        # Y = [rand(Bernoulli(p)) for p in ps]
-        return X, Y
-    end
-
-    dist_y_given_f(f) = Bernoulli(logistic(f))
-
-    function build_latent_gp(theta)
-        variance = softplus(theta[1])
-        lengthscale = softplus(theta[2])
-        kernel = variance * with_lengthscale(SqExponentialKernel(), lengthscale)
-        return LatentGP(GP(kernel), dist_y_given_f, 1e-8)
-    end
+    generate_data = ApproximateGPs.TestUtils.generate_data
+    dist_y_given_f = ApproximateGPs.TestUtils.dist_y_given_f
+    build_latent_gp = ApproximateGPs.TestUtils.build_latent_gp
 
     function optimize_elbo(
         build_latent_gp,
@@ -49,43 +28,8 @@
     end
 
     @testset "predictions" begin
-        rng = MersenneTwister(123456)
-        N_cond = 5
-        N_a = 6
-        N_b = 7
-
-        # Specify prior.
-        f = GP(Matern32Kernel())
-        # Sample from prior.
-        x = collect(range(-1.0, 1.0; length=N_cond))
-        noise_scale = 0.1
-        fx = f(x, noise_scale^2)
-        y = rand(rng, fx)
-
-        jitter = 0.0  # not needed in Gaussian case
-        lf = LatentGP(f, f -> Normal(f, noise_scale), jitter)
-        # in Gaussian case, Laplace converges to f_opt in one step; we need the
-        # second step to compute the cache at f_opt rather than f_init!
-        f_approx_post = posterior(LaplaceApproximation(; maxiter=2), lf(x), y)
-
-        @testset "AbstractGPs API" begin
-            a = collect(range(-1.2, 1.2; length=N_a))
-            b = randn(rng, N_b)
-            AbstractGPs.TestUtils.test_internal_abstractgps_interface(
-                rng, f_approx_post, a, b
-            )
-        end
-
-        @testset "equivalence to exact GPR for Gaussian likelihood" begin
-            f_exact_post = posterior(f(x, noise_scale^2), y)
-            xt = vcat(x, randn(rng, 3))  # test at training and new points
-
-            m_approx, c_approx = mean_and_cov(f_approx_post(xt))
-            m_exact, c_exact = mean_and_cov(f_exact_post(xt))
-
-            @test m_approx ≈ m_exact
-            @test c_approx ≈ c_exact
-        end
+        approx = LaplaceApproximation(; maxiter=2)
+        ApproximateGPs.TestUtils.test_approximation_predictions(approx)
     end
 
     @testset "gradients" begin

test/runtests.jl

@@ -1,10 +1,8 @@
 using Random
 using Test
-using ApproximateGPs
 using Flux
 using IterTools
 using AbstractGPs
-using AbstractGPs: LatentFiniteGP, TestUtils
 using Distributions
 using LogExpFunctions: logistic
 using LinearAlgebra
@@ -14,6 +12,8 @@ using Zygote
 using ChainRulesCore
 using ChainRulesTestUtils
 using FiniteDifferences
+
+using ApproximateGPs
 using ApproximateGPs: SparseVariationalApproximationModule, LaplaceApproximationModule
 
 # Writing tests: