test: compile functions in tests correctly

LuxDL · Nov 14, 2024 · 5dd9c16 · 5dd9c16
1 parent b1aee66
commit 5dd9c16
Show file tree

Hide file tree

Showing 2 changed files with 18 additions and 13 deletions.
diff --git a/test/reactant/loss_tests.jl b/test/reactant/loss_tests.jl
@@ -72,10 +72,12 @@
                 @test celoss_ls(ŷ, y) ≈ @jit(celoss_ls(ŷ_ra, y_ra))
 
                 celoss_lp = CrossEntropyLoss(; logits=Val(true))
-                @test celoss_lp(log.(ŷ), y) ≈ @jit(celoss_lp(log.(ŷ_ra), y_ra))
+                logit_celoss_lp = (ŷ, y) -> celoss_lp(log.(ŷ), y)
+                @test logit_celoss_lp(ŷ, y) ≈ @jit(logit_celoss_lp(ŷ_ra, y_ra))
 
                 celoss_lp_ls = CrossEntropyLoss(; logits=Val(true), label_smoothing=0.1)
-                @test celoss_lp_ls(log.(ŷ), y) ≈ @jit(celoss_lp_ls(log.(ŷ_ra), y_ra))
+                logit_celoss_lp_ls = (ŷ, y) -> celoss_lp_ls(log.(ŷ), y)
+                @test logit_celoss_lp_ls(ŷ, y) ≈ @jit(logit_celoss_lp_ls(ŷ_ra, y_ra))
             end
 
             @testset "Binary CrossEntropyLoss" begin
@@ -86,11 +88,13 @@
                 @test bceloss_ls(ŷ, y) ≈ @jit(bceloss_ls(ŷ_ra, y_ra))
 
                 bceloss_lp = BinaryCrossEntropyLoss(; logits=Val(true))
-                @test bceloss_lp(log.(ŷ), y) ≈ @jit(bceloss_lp(log.(ŷ_ra), y_ra))
+                logit_bceloss_lp = (ŷ, y) -> bceloss_lp(log.(ŷ), y)
+                @test logit_bceloss_lp(ŷ, y) ≈ @jit(logit_bceloss_lp(ŷ_ra, y_ra))
 
                 bceloss_lp_ls = BinaryCrossEntropyLoss(;
                     logits=Val(true), label_smoothing=0.1)
-                @test bceloss_lp_ls(log.(ŷ), y) ≈ @jit(bceloss_lp_ls(log.(ŷ_ra), y_ra))
+                logit_bceloss_lp_ls = (ŷ, y) -> bceloss_lp_ls(log.(ŷ), y)
+                @test logit_bceloss_lp_ls(ŷ, y) ≈ @jit(logit_bceloss_lp_ls(ŷ_ra, y_ra))
             end
 
             @testset "BinaryFocalLoss" begin

diff --git a/test/reactant/training_tests.jl b/test/reactant/training_tests.jl
@@ -24,18 +24,22 @@
             ps, st = Lux.setup(StableRNG(1234), model) |> xdev
 
             x_ra = randn(Float32, 2, 32) |> xdev
+            y_ra = rand(Float32, 2, 32) |> xdev
 
-            inference_fn = @compile model(x_ra, ps, Lux.testmode(st))
+            inference_loss_fn = (xᵢ, yᵢ, mode, ps, st) -> begin
+                ŷᵢ, _ = model(xᵢ, ps, Lux.testmode(st))
+                return MSELoss()(ŷᵢ, yᵢ)
+            end
+            inference_loss_fn_compiled = @compile inference_loss_fn(
+                x_ra, y_ra, model, ps, st
+            )
 
             x = [rand(Float32, 2, 32) for _ in 1:32]
             y = [xᵢ .^ 2 for xᵢ in x]
 
             dataloader = DeviceIterator(xdev, zip(x, y))
 
-            total_initial_loss = mapreduce(+, dataloader) do (xᵢ, yᵢ)
-                ŷᵢ, _ = inference_fn(xᵢ, ps, Lux.testmode(st))
-                return MSELoss()(ŷᵢ, yᵢ)
-            end
+            total_initial_loss = mapreduce(inference_loss_fn_compiled, +, dataloader)
 
             train_state = Training.TrainState(model, ps, st, Adam(0.01f0))
 
@@ -51,10 +55,7 @@
                 end
             end
 
-            total_final_loss = mapreduce(+, dataloader) do (xᵢ, yᵢ)
-                ŷᵢ, _ = inference_fn(xᵢ, train_state.parameters, Lux.testmode(st))
-                return MSELoss()(ŷᵢ, yᵢ)
-            end
+            total_final_loss = mapreduce(inference_loss_fn_compiled, +, dataloader)
 
             @test total_final_loss < 100 * total_initial_loss
         end