Modularize package tests (#17)

* Modularize tests

* Fix type inferrability

* Fix Flux GPU URL

Author: adrhill

docs/src/literate/basics.jl

@@ -142,7 +142,7 @@ expl.extras.layerwise_relevances
 # ## Performance tips
 # ### Using LRP with a GPU
 # All LRP analyzers support GPU backends,
-# building on top of [Flux.jl's GPU support](https://fluxml.ai/Flux.jl/stable/gpu/).
+# building on top of [Flux.jl's GPU support](https://fluxml.ai/Flux.jl/stable/guide/gpu/).
 # Using a GPU only requires moving the input array and model weights to the GPU.
 #
 # For example, using [CUDA.jl](https://github.com/JuliaGPU/CUDA.jl):

src/rules.jl

@@ -16,6 +16,7 @@ function lrp!(Rᵏ, rule::AbstractLRPRule, layer, modified_layer, aᵏ, Rᵏ⁺
     s = Rᵏ⁺¹ ./ modify_denominator(rule, z)
     c = only(back(s))
     Rᵏ .= ãᵏ .* c
+    return Rᵏ
 end
 
 #===================================#

test/Project.toml

@@ -11,4 +11,3 @@ ReferenceTests = "324d217c-45ce-50fc-942e-d289b448e8cf"
 Suppressor = "fd094767-a336-5f1f-9728-57cf17d0bbfb"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 XAIBase = "9b48221d-a747-4c1b-9860-46a1d8ba24a7"
-

test/runtests.jl

@@ -1,24 +1,23 @@
-using Test
-using ReferenceTests
-using Aqua
-using JuliaFormatter
-using Random
-
 using RelevancePropagation
-using Flux
-import Flux: Scale
 
-pseudorand(dims...) = rand(MersenneTwister(123), Float32, dims...)
+using Test
+using JuliaFormatter
+using Aqua
 
 @testset "RelevancePropagation.jl" begin
-    @testset "Aqua.jl" begin
-        @info "Running Aqua.jl's auto quality assurance tests. These might print warnings from dependencies."
-        Aqua.test_all(RelevancePropagation; ambiguities=false)
-    end
-    @testset "JuliaFormatter.jl" begin
-        @info "Running JuliaFormatter's code formatting tests."
-        @test format(RelevancePropagation; verbose=false, overwrite=false)
+    if VERSION >= v"1.10"
+        @testset "Code formatting" begin
+            @info "- Testing code formatting with JuliaFormatter..."
+            @test JuliaFormatter.format(
+                RelevancePropagation; verbose=false, overwrite=false
+            )
+        end
+        @testset "Aqua.jl" begin
+            @info "- Running Aqua.jl tests. These might print warnings from dependencies..."
+            Aqua.test_all(RelevancePropagation; ambiguities=false)
+        end
     end
+
     @testset "Utilities" begin
         @info "Testing utilities..."
         include("test_utils.jl")

test/test_batches.jl

@@ -1,6 +1,10 @@
-using Flux
 using RelevancePropagation
-using Random
+using Test
+
+using Flux
+using Random: rand, MersenneTwister
+
+pseudorand(dims...) = rand(MersenneTwister(123), Float32, dims...)
 
 ## Test `fuse_batchnorm` on Dense and Conv layers
 ins = 20

test/test_canonize.jl

@@ -1,9 +1,13 @@
+using RelevancePropagation
+using Test
+
 using Flux
 using Flux: flatten, Scale
-using RelevancePropagation
 using RelevancePropagation: canonize_fuse
 using Random
 
+pseudorand(dims...) = rand(MersenneTwister(123), Float32, dims...)
+
 batchsize = 50
 
 ##=====================================#

test/test_chain_utils.jl

@@ -1,3 +1,6 @@
+using RelevancePropagation
+using Test
+
 using RelevancePropagation: ChainTuple, ParallelTuple, SkipConnectionTuple
 using RelevancePropagation: ModelIndex, chainmap, chainindices, chainzip
 using RelevancePropagation: activation_fn

test/test_checks.jl

@@ -1,6 +1,10 @@
 using RelevancePropagation
+using Test
+using ReferenceTests
+
 using RelevancePropagation: check_lrp_compat, print_lrp_model_check
 using Suppressor
+
 err = ErrorException("Unknown layer or activation function found in model")
 
 # Flux layers

test/test_cnn.jl

@@ -1,13 +1,19 @@
 using RelevancePropagation
+using Test
+using ReferenceTests
+
 using Flux
 using JLD2
+using Random: rand, MersenneTwister
 
 const LRP_ANALYZERS = Dict(
     "LRPZero"                   => LRP,
     "LRPZero_COC"               => m -> LRP(m; flatten=false), # chain of chains
     "LRPEpsilonAlpha2Beta1Flat" => m -> LRP(m, EpsilonAlpha2Beta1Flat()),
 )
 
+pseudorand(dims...) = rand(MersenneTwister(123), Float32, dims...)
+
 input_size = (32, 32, 3, 1)
 input = pseudorand(input_size)
 

test/test_composite.jl

@@ -1,6 +1,11 @@
 using RelevancePropagation
+using Test
+using ReferenceTests
+
+using NNlib
+using Flux
+using Flux: flatten, Scale
 using Metalhead
-using Flux, NNlib
 
 model = VGG(11; pretrain=false).layers
 model_flat = flatten_model(model)