Allow passing rng to CV fold construction

src/GLMNet.jl

@@ -505,10 +505,11 @@ end
 
 function glmnetcv(X::AbstractMatrix, y::Union{AbstractVector,AbstractMatrix},
                   family::Distribution=Normal(); weights::Vector{Float64}=ones(length(y)),
+                  rng=Random.GLOBAL_RNG,
                   nfolds::Int=min(10, div(size(y, 1), 3)),
                   folds::Vector{Int}=begin
                       n, r = divrem(size(y, 1), nfolds)
-                      shuffle!([repeat(1:nfolds, outer=n); 1:r])
+                      shuffle!(rng, [repeat(1:nfolds, outer=n); 1:r])
                   end, parallel::Bool=false, kw...)
     # Fit full model once to determine parameters
     X = convert(Matrix{Float64}, X)

test/runtests.jl

@@ -1,5 +1,5 @@
 using GLMNet, Distributions
-using SparseArrays, Test
+using Random, SparseArrays, Test
 
 X = [74    1  93  93  79  18
      98   36   2  27  65  70
@@ -146,6 +146,16 @@ cv = glmnetcv(X, y; folds=[1,1,1,1,2,2,2,3,3,3])
 show(IOBuffer(), cv)
 show(IOBuffer(), cv.path)
 
+# Passing RNG makes cv deterministic
+cv1 = glmnetcv(X, y)
+cv2 = glmnetcv(X, y)
+@test cv1.meanloss ≉ cv2.meanloss
+@test cv1.stdloss ≉ cv2.stdloss
+cv3 = glmnetcv(X, y; rng=MersenneTwister(1))
+cv4 = glmnetcv(X, y; rng=MersenneTwister(1))
+@test cv3.meanloss ≈ cv4.meanloss
+@test cv3.stdloss ≈ cv4.stdloss
+
 ## LOGISTIC
 yl = [(y .< 50) (y .>= 50)]
 df_true = [0,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
@@ -302,6 +312,16 @@ cv = glmnetcv(X, yl, Binomial(); folds=[1,1,1,1,2,2,2,3,3,3])
 show(IOBuffer(), cv)
 show(IOBuffer(), cv.path)
 
+# Passing RNG makes cv deterministic
+cv1 = glmnetcv(X, yl, Binomial())
+cv2 = glmnetcv(X, yl, Binomial())
+@test cv1.meanloss ≉ cv2.meanloss
+@test cv1.stdloss ≉ cv2.stdloss
+cv3 = glmnetcv(X, yl, Binomial(); rng=MersenneTwister(1))
+cv4 = glmnetcv(X, yl, Binomial(); rng=MersenneTwister(1))
+@test cv3.meanloss ≈ cv4.meanloss
+@test cv3.stdloss ≈ cv4.stdloss
+
 # Make sure passing nlambda to glmnetcv works
 cv = glmnetcv(X, y, Poisson(); nlambda=2, lambda_min_ratio=0.01)
 @test length(cv.lambda) == 2
@@ -462,6 +482,17 @@ cv = glmnetcv(X, y, Poisson(); folds=[1,1,1,1,2,2,2,3,3,3])
 show(IOBuffer(), cv)
 show(IOBuffer(), cv.path)
 
+# Passing RNG makes cv deterministic
+cv1 = glmnetcv(X, y, Poisson())
+cv2 = glmnetcv(X, y, Poisson())
+@test cv1.meanloss ≉ cv2.meanloss
+@test cv1.stdloss ≉ cv2.stdloss
+cv3 = glmnetcv(X, y, Poisson(); rng=MersenneTwister(1))
+cv4 = glmnetcv(X, y, Poisson(); rng=MersenneTwister(1))
+@test cv3.meanloss ≈ cv4.meanloss
+@test cv3.stdloss ≈ cv4.stdloss
+
+
 ## COMPRESSEDPREDICTORMATRIX
 betas = path.betas
 cbetas = convert(Matrix{Float64}, path.betas)