Switch to StableRNGs for broken tests.

Author: yebai

test/Project.toml

@@ -27,6 +27,7 @@ StatsFuns = "4c63d2b9-4356-54db-8cca-17b64c39e42c"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 Tracker = "9f7883ad-71c0-57eb-9f7f-b5c9e6d3789c"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
+StableRNGs = "860ef19b-820b-49d6-a774-d7a799459cd3"
 
 [compat]
 AbstractMCMC = "3.2.1"
@@ -53,4 +54,5 @@ StatsBase = "0.33"
 StatsFuns = "0.9.5"
 Tracker = "0.2.11"
 Zygote = "0.5.4, 0.6"
+StableRNGs = "1"
 julia = "1.3"

test/contrib/inference/sghmc.jl

@@ -16,10 +16,10 @@
         @test sampler isa Turing.Sampler{<:SGHMC}
     end
     @numerical_testset "sghmc inference" begin
-        Random.seed!(54321)
+        rng = StableRNG(123)
 
         alg = SGHMC(; learning_rate=0.02, momentum_decay=0.5)
-        chain = sample(gdemo_default, alg, 10_000)
+        chain = sample(rng, gdemo_default, alg, 10_000)
         check_gdemo(chain, atol = 0.1)
     end
 end
@@ -42,9 +42,9 @@ end
         @test sampler isa Turing.Sampler{<:SGLD}
     end
     @numerical_testset "sgld inference" begin
-        Random.seed!(12345)
+        rng = StableRNG(1)
 
-        chain = sample(gdemo_default, SGLD(; stepsize = PolynomialStepsize(0.5)), 10_000)
+        chain = sample(rng, gdemo_default, SGLD(; stepsize = PolynomialStepsize(0.5)), 20_000)
         check_gdemo(chain, atol = 0.2)
 
         # Weight samples by step sizes (cf section 4.2 in the paper by Welling and Teh)

test/inference/hmc.jl

@@ -1,8 +1,7 @@
 @testset "hmc.jl" begin
+    # Set a seed
+    rng = StableRNG(123)
     @numerical_testset "constrained bounded" begin
-        # Set a seed
-        Random.seed!(5)
-        
         obs = [0,1,0,1,1,1,1,1,1,1]
 
         @model constrained_test(obs) = begin
@@ -14,6 +13,7 @@
         end
 
         chain = sample(
+            rng,
             constrained_test(obs),
             HMC(1.5, 3),# using a large step size (1.5)
             1000)
@@ -33,16 +33,16 @@
         end
 
         chain = sample(
+            rng,
             constrained_simplex_test(obs12),
             HMC(0.75, 2),
             1000)
 
         check_numerical(chain, ["ps[1]", "ps[2]"], [5/16, 11/16], atol=0.015)
     end
     @numerical_testset "hmc reverse diff" begin
-        Random.seed!(1)
         alg = HMC(0.1, 10)
-        res = sample(gdemo_default, alg, 4000)
+        res = sample(rng, gdemo_default, alg, 4000)
         check_gdemo(res, rtol=0.1)
     end
     @turing_testset "matrix support" begin
@@ -53,7 +53,7 @@
         model_f = hmcmatrixsup()
         n_samples = 1_000
         vs = map(1:3) do _
-            chain = sample(model_f, HMC(0.15, 7), n_samples)
+            chain = sample(rng, model_f, HMC(0.15, 7), n_samples)
             r = reshape(Array(group(chain, :v)), n_samples, 2, 2)
             reshape(mean(r; dims = 1), 2, 2)
         end
@@ -103,25 +103,24 @@
         end
 
         # Sampling
-        chain = sample(bnn(ts), HMC(0.1, 5), 10)
+        chain = sample(rng, bnn(ts), HMC(0.1, 5), 10)
     end
     @numerical_testset "hmcda inference" begin 
-        Random.seed!(12345)
         alg1 = HMCDA(1000, 0.8, 0.015)
         # alg2 = Gibbs(HMCDA(200, 0.8, 0.35, :m), HMC(0.25, 3, :s))
         alg3 = Gibbs(PG(10, :s), HMCDA(200, 0.8, 0.005, :m))
         # alg3 = Gibbs(HMC(0.25, 3, :m), PG(30, 3, :s))
         # alg3 = PG(50, 2000)
 
-        res1 = sample(gdemo_default, alg1, 3000)
+        res1 = sample(rng, gdemo_default, alg1, 3000)
         check_gdemo(res1)
 
         # res2 = sample(gdemo([1.5, 2.0]), alg2)
         #
         # @test mean(res2[:s]) ≈ 49/24 atol=0.2
         # @test mean(res2[:m]) ≈ 7/6 atol=0.2
 
-        res3 = sample(gdemo_default, alg3, 2000)
+        res3 = sample(rng, gdemo_default, alg3, 2000)
         check_gdemo(res3)
     end
 
@@ -146,7 +145,7 @@
     end
     @numerical_testset "nuts inference" begin
         alg = NUTS(1000, 0.8)
-        res = sample(gdemo_default, alg, 6000)
+        res = sample(rng, gdemo_default, alg, 6000)
         check_gdemo(res)
     end
     @turing_testset "nuts constructor" begin
@@ -165,8 +164,8 @@
     @turing_testset "check discard" begin
         alg = NUTS(100, 0.8)
 
-        c1 = sample(gdemo_default, alg, 500, discard_adapt = true)
-        c2 = sample(gdemo_default, alg, 500, discard_adapt = false)
+        c1 = sample(rng, gdemo_default, alg, 500, discard_adapt = true)
+        c2 = sample(rng, gdemo_default, alg, 500, discard_adapt = false)
 
         @test size(c1, 1) == 500
         @test size(c2, 1) == 500
@@ -175,9 +174,9 @@
         alg1 = Gibbs(PG(10, :m), NUTS(100, 0.65, :s))
         alg2 = Gibbs(PG(10, :m), HMC(0.1, 3, :s))
         alg3 = Gibbs(PG(10, :m), HMCDA(100, 0.65, 0.3, :s))
-        @test sample(gdemo_default, alg1, 300) isa Chains
-        @test sample(gdemo_default, alg2, 300) isa Chains
-        @test sample(gdemo_default, alg3, 300) isa Chains
+        @test sample(rng, gdemo_default, alg1, 300) isa Chains
+        @test sample(rng, gdemo_default, alg2, 300) isa Chains
+        @test sample(rng, gdemo_default, alg3, 300) isa Chains
     end
 
     @turing_testset "Regression tests" begin
@@ -186,19 +185,19 @@
             m = Matrix{T}(undef, 2, 3)
             m .~ MvNormal(zeros(2), I)
         end
-        @test sample(mwe1(), HMC(0.2, 4), 1_000) isa Chains
+        @test sample(rng, mwe1(), HMC(0.2, 4), 1_000) isa Chains
 
         @model function mwe2(::Type{T} = Matrix{Float64}) where T
             m = T(undef, 2, 3)
             m .~ MvNormal(zeros(2), I)
         end
-        @test sample(mwe2(), HMC(0.2, 4), 1_000) isa Chains
+        @test sample(rng, mwe2(), HMC(0.2, 4), 1_000) isa Chains
 
         # https://github.com/TuringLang/Turing.jl/issues/1308
         @model function mwe3(::Type{T} = Array{Float64}) where T
             m = T(undef, 2, 3)
             m .~ MvNormal(zeros(2), I)
         end
-        @test sample(mwe3(), HMC(0.2, 4), 1_000) isa Chains
+        @test sample(rng, mwe3(), HMC(0.2, 4), 1_000) isa Chains
     end
 end

test/runtests.jl

@@ -26,6 +26,7 @@ using LinearAlgebra
 using Pkg
 using Random
 using Test
+using StableRNGs
 
 using AdvancedPS: ResampleWithESSThreshold, resample_systematic, resample_multinomial
 using AdvancedVI: TruncatedADAGrad, DecayedADAGrad, apply!

test/stdlib/distributions.jl

@@ -1,20 +1,19 @@
 @testset "distributions.jl" begin
+    rng = StableRNG(12345)
     @turing_testset "distributions functions" begin
         ns = 10
-        logitp = randn()
+        logitp = randn(rng)
         d1 = BinomialLogit(ns, logitp)
         d2 = Binomial(ns, logistic(logitp))
         k = 3
         @test logpdf(d1, k) ≈ logpdf(d2, k)
     end
 
     @turing_testset "distributions functions" begin
-        Random.seed!(1)
-
         d = OrderedLogistic(-2, [-1, 1])
 
         n = 1_000_000
-        y = rand(d, n)
+        y = rand(rng, d, n)
         K = length(d.cutpoints) + 1
         p = [mean(==(k), y) for k in 1:K]          # empirical probs
         pmf = [exp(logpdf(d, k)) for k in 1:K]
@@ -31,9 +30,9 @@
     end
 
     @numerical_testset "single distribution correctness" begin
-        Random.seed!(12345)
+        rng = StableRNG(1)
 
-        n_samples = 5_000
+        n_samples = 10_000
         mean_tol = 0.1
         var_atol = 1.0
         var_tol = 0.5
@@ -113,7 +112,7 @@
 
                         @model m() = x ~ dist
 
-                        chn = sample(m(), HMC(0.05, 20), n_samples)
+                        chn = sample(rng, m(), HMC(0.05, 20), n_samples)
 
                         # Numerical tests.
                         check_dist_numerical(dist,