use default_rng when comparing acceleration methods

FluxML · May 1, 2024 · 99100b8 · 99100b8
1 parent c806dc5
commit 99100b8
Show file tree

Hide file tree

Showing 4 changed files with 6 additions and 8 deletions.
diff --git a/test/classifier.jl b/test/classifier.jl
@@ -62,9 +62,8 @@ losses = []
     # check flux model is an improvement on predicting constant
     # distribution
     # (GPUs only support `default_rng`):
-    rng = accel == CPU1() ? StableRNGs.StableRNG(123) : Random.default_rng()
+    rng = Random.default_rng()
     seed!(rng, 123)
-    rng = StableRNGs.StableRNG(123)
     model = MLJFlux.NeuralNetworkClassifier(epochs=50,
                                             builder=builder,
                                             optimiser=optimiser,

diff --git a/test/image.jl b/test/image.jl
@@ -33,7 +33,7 @@ losses = []
 @testset_accelerated "ImageClassifier basic tests" accel begin
 
     # GPUs only support `default_rng`:
-    rng = accel == CPU1() ? StableRNGs.StableRNG(123) : Random.default_rng()
+    rng = Random.default_rng()
     seed!(rng, 123)
 
     model = MLJFlux.ImageClassifier(builder=builder,
@@ -86,7 +86,7 @@ losses = []
 @testset_accelerated "ColorImages" accel begin
 
     # GPUs only support `default_rng`:
-    rng = accel == CPU1() ? StableRNGs.StableRNG(123) : Random.default_rng()
+    rng = Random.default_rng()
     seed!(rng, 123)
 
     model = MLJFlux.ImageClassifier(builder=builder,
@@ -129,7 +129,7 @@ noise=0.2, color=true);
 @testset_accelerated "ImageClassifier basic tests" accel begin
 
     # GPUs only support `default_rng`:
-    rng = accel == CPU1() ? StableRNGs.StableRNG(123) : Random.default_rng()
+    rng = Random.default_rng()
     seed!(rng, 123)
 
     model = MLJFlux.ImageClassifier(epochs=5,

diff --git a/test/regressor.jl b/test/regressor.jl
@@ -42,7 +42,7 @@ train, test = MLJBase.partition(1:N, 0.7)
 
     # test model is a bit better than constant predictor:
     # (GPUs only support `default_rng`):
-    rng = accel == CPU1() ? StableRNGs.StableRNG(123) : Random.default_rng()
+    rng = Random.default_rng()
     seed!(rng, 123)
     model = MLJFlux.NeuralNetworkRegressor(builder=builder,
                                            acceleration=accel,
@@ -108,7 +108,7 @@ losses = []
 
     # test model is a bit better than constant predictor
     # (GPUs only support `default_rng`):
-    rng = accel == CPU1() ? StableRNGs.StableRNG(123) : Random.default_rng()
+    rng = Random.default_rng()
     seed!(rng, 123)
     model = MLJFlux.MultitargetNeuralNetworkRegressor(
         acceleration=accel,

diff --git a/test/test_utils.jl b/test/test_utils.jl
@@ -164,7 +164,6 @@ function optimisertest(ModelType, X, y, builder, optimiser, accel)
                  @test isapprox(l1, l2)
              else
                  @test_broken isapprox(l1, l2, rtol=1e-8)
-                 @show l1/l2
              end
 
              # USING USER SPECIFIED RNG SEED (unsupported on GPU)