Add more tests for single distributions (#621)

* add uni tests

* force string conversion

* add more dists

* fix typo in runtest

* exclude var for MatrixDist

* improve script

* set a upper limit for the stepsize in adaptation

* cap end-of-window step size update

* improve type stability

* resolve Mohamed's comment

* add inv wishart test (suggested by Will)

* add more continous univariate dists

* fix incorrect normal inverse gaussian parameters

* add warning msg for capping ss to 1.0

* comment out VonMises distribution

src/inference/adapt/adapt.jl

@@ -85,7 +85,8 @@ function adapt!(tp::ThreePhaseAdapter, stats::Real, θ; adapt_ϵ=false, adapt_M=
         tp.state.n += 1
         if tp.state.n == tp.n_adapts
             if adapt_ϵ
-                tp.ssa.state.ϵ = exp(tp.ssa.state.x_bar)
+                ϵ = exp(tp.ssa.state.x_bar)
+                tp.ssa.state.ϵ = min(one(ϵ), ϵ)
             end
             @info " Adapted ϵ = $(getss(tp)), std = $(string(tp.pc)); $(tp.state.n) iterations is used for adaption."
         else

src/inference/adapt/stepsize.jl

@@ -93,7 +93,8 @@ function adapt_stepsize!(da::DualAveraging, stats::Real)
     if isnan(ϵ) || isinf(ϵ)
         @warn "Incorrect ϵ = $ϵ; ϵ_previous = $(da.state.ϵ) is used instead."
     else
-        da.state.ϵ = ϵ
+        ϵ < one(ϵ) && @warn "$ϵ exceeds 1.0; capped to 1.0 for numerical stability"
+        da.state.ϵ = min(one(ϵ), ϵ)
     end
     da.state.x_bar = x_bar
     da.state.H_bar = H_bar

test/models.jl/single_dist_correctness.jl

@@ -0,0 +1,105 @@
+using Test, Turing
+turnprogress(false)
+
+n_samples = 20_000
+mean_atol = 0.25
+var_atol = 1.0
+multi_dim = 10
+
+# 1. UnivariateDistribution
+# NOTE: Noncentral distributions are commented out because of
+#       AD imcompatibility of their logpdf functions
+dist_uni = [Arcsine(1, 3),
+            Beta(2, 1),
+            # NoncentralBeta(2, 1, 1),
+            BetaPrime(1, 1),
+            Biweight(0, 1),
+            Chi(7),
+            Chisq(7),
+            # NoncentralChisq(7, 1),
+            Cosine(0, 1),
+            Epanechnikov(0, 1),
+            Erlang(2, 3),
+            Exponential(0.1),
+            FDist(7, 7),
+            # NoncentralF(7, 7, 1),
+            Frechet(2, 0.5),
+            Normal(0, 1),
+            GeneralizedExtremeValue(0, 1, 0.5),
+            GeneralizedPareto(0, 1, 0.5),
+            Gumbel(0, 0.5),
+            InverseGaussian(1, 1),
+            Kolmogorov(),
+            # KSDist(2),  # no pdf function defined
+            # KSOneSided(2),  # no pdf function defined
+            Laplace(0, 0.5),
+            Levy(0, 1),
+            Logistic(0, 1),
+            LogNormal(0, 1),
+            Gamma(2, 3),
+            InverseGamma(3, 1),
+            NormalCanon(0, 1),
+            NormalInverseGaussian(0, 2, 1, 1),
+            Pareto(1, 1),
+            Rayleigh(1),
+            SymTriangularDist(0, 1),
+            TDist(2.5),
+            # NoncentralT(2.5, 1),
+            TriangularDist(1, 3, 2),
+            Triweight(0, 1),
+            Uniform(0, 1),
+            # VonMises(0, 1), WARNING: this is commented are because the test is broken
+            Weibull(2, 1),
+            # Cauchy(0, 1),  # mean and variance are undefined for Cauchy
+           ]
+
+# 2. MultivariateDistribution
+dist_multi = [MvNormal(zeros(multi_dim), ones(multi_dim)),
+              MvNormal(zeros(2), [2 1; 1 4]),
+              Dirichlet(multi_dim, 2.0),
+             ]
+
+# 3. MatrixDistribution
+dist_matrix = [Wishart(7, [1 0.5; 0.5 1]),
+               InverseWishart(7, [1 0.5; 0.5 1]),
+              ]
+
+@testset "Correctness test for single distributions" begin
+    for (dist_set, dist_list) ∈ [("UnivariateDistribution",   dist_uni),
+                                 ("MultivariateDistribution", dist_multi),
+                                 ("MatrixDistribution",       dist_matrix)
+                                 ]
+        @testset "$(string(dist_set))" begin
+            for dist in dist_list
+                @testset "$(string(typeof(dist)))" begin
+                    @info "Distribution(params)" dist
+
+                    @model m() = begin
+                        x ~ dist
+                    end
+                    chn = sample(m(), NUTS(n_samples, 0.8))
+
+                    chn_xs = chn[:x][1:2:end] # thining by halving
+
+                    # Mean
+                    dist_mean = mean(dist)
+                    if !all(isnan.(dist_mean)) && !all(isinf.(dist_mean))
+                        chn_mean = mean(chn_xs)
+                        @test chn_mean ≈ dist_mean atol=(mean_atol * length(chn_mean))
+                    end
+
+                    # var() for Distributions.MatrixDistribution is not defined
+                    if !(dist isa Distributions.MatrixDistribution)
+                        # Variance
+                        dist_var = var(dist)
+                        if !all(isnan.(dist_var)) && !all(isinf.(dist_var))
+                            chn_var = var(chn_xs)
+                            @test chn_var ≈ dist_var atol=(var_atol * length(chn_var))
+                        end
+                    end
+                end
+            end
+        end
+    end
+end
+# Wishart(7, [1 0.5; 0.5 1])

test/runtests.jl

@@ -14,7 +14,8 @@ cd(path); include("utility.jl")
 if get(ENV, "TRAVIS", "false") == "true"
     # If Travis is testing, separate the tests.
     numerical_tests = [joinpath("hmc.jl", "matrix_support.jl"),
-                       joinpath("mh.jl", "mh_cons.jl")]
+                       joinpath("mh.jl", "mh_cons.jl"),
+                       joinpath("models.jl", "single_dist_correctness.jl")]
 
     if ENV["STAGE"] == "test"
         runtests(exclude = numerical_tests)