Add GradCAM analyzer

README.md

@@ -101,6 +101,7 @@ Currently, the following analyzers are implemented:
 * `InputTimesGradient`
 * `SmoothGrad`
 * `IntegratedGradients`
+* `GradCAM`
 
 One of the design goals of the [Julia-XAI ecosystem][juliaxai-docs] is extensibility.
 To implement an XAI method, take a look at the [common interface

docs/src/api.md

@@ -12,6 +12,7 @@ Gradient
 InputTimesGradient
 SmoothGrad
 IntegratedGradients
+GradCAM
 ```
 
 # Input augmentations

src/ExplainableAI.jl

@@ -12,9 +12,11 @@ include("compat.jl")
 include("bibliography.jl")
 include("input_augmentation.jl")
 include("gradient.jl")
+include("gradcam.jl")
 
 export Gradient, InputTimesGradient
 export NoiseAugmentation, SmoothGrad
 export InterpolationAugmentation, IntegratedGradients
+export GradCAM
 
 end # module

src/bibliography.jl

@@ -1,3 +1,4 @@
 # Gradient methods:
 const REF_SMILKOV_SMOOTHGRAD = "Smilkov et al., *SmoothGrad: removing noise by adding noise*"
 const REF_SUNDARARAJAN_AXIOMATIC = "Sundararajan et al., *Axiomatic Attribution for Deep Networks*"
+const REF_SELVARAJU_GRADCAM = "Selvaraju et al., *Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization*"

src/gradcam.jl

@@ -0,0 +1,31 @@
+"""
+    GradCAM(feature_layers, adaptation_layers)
+
+Calculates the Gradient-weighted Class Activation Map (GradCAM).
+GradCAM provides a visual explanation of the regions with significant neuron importance for the model's classification decision.
+
+# Parameters
+- `feature_layers`: The layers of a convolutional neural network (CNN) responsible for extracting feature maps.
+- `adaptation_layers`: The layers of the CNN used for adaptation and classification.
+
+# Note
+Flux is not required for GradCAM. 
+GradCAM is compatible with a wide variety of CNN model-families.
+
+# References
+- $REF_SELVARAJU_GRADCAM
+"""
+struct GradCAM{F,A} <: AbstractXAIMethod
+    feature_layers::F
+    adaptation_layers::A
+end
+function (analyzer::GradCAM)(input, ns::AbstractNeuronSelector)
+    A = analyzer.feature_layers(input)  # feature map
+    feature_map_size = size(A, 1) * size(A, 2)
+
+    # Determine neuron importance αₖᶜ = 1/Z * ∑ᵢ ∑ⱼ ∂yᶜ / ∂Aᵢⱼᵏ
+    grad, output, output_indices = gradient_wrt_input(analyzer.adaptation_layers, A, ns)
+    αᶜ = sum(grad; dims=(1, 2)) / feature_map_size
+    Lᶜ = max.(sum(αᶜ .* A; dims=3), 0)
+    return Explanation(Lᶜ, output, output_indices, :GradCAM, :cam, nothing)
+end

test/test_batches.jl

@@ -8,7 +8,7 @@ ins = 20
 outs = 10
 batchsize = 15
 
-model = Chain(Dense(ins, outs, relu; init=pseudorand))
+model = Chain(Dense(ins, 15, relu; init=pseudorand), Dense(15, outs, relu; init=pseudorand))
 
 # Input 1 w/o batch dimension
 input1_no_bd = rand(MersenneTwister(1), Float32, ins)
@@ -24,6 +24,7 @@ ANALYZERS = Dict(
     "InputTimesGradient"  => InputTimesGradient,
     "SmoothGrad"          => m -> SmoothGrad(m, 5, 0.1, MersenneTwister(123)),
     "IntegratedGradients" => m -> IntegratedGradients(m, 5),
+    "GradCAM"             => m -> GradCAM(m[1], m[2]),
 )
 
 for (name, method) in ANALYZERS

test/test_cnn.jl

@@ -6,6 +6,7 @@ const GRADIENT_ANALYZERS = Dict(
     "InputTimesGradient"  => InputTimesGradient,
     "SmoothGrad"          => m -> SmoothGrad(m, 5, 0.1, MersenneTwister(123)),
     "IntegratedGradients" => m -> IntegratedGradients(m, 5),
+    "GradCAM"             => m -> GradCAM(m[1], m[2]),
 )
 
 input_size = (32, 32, 3, 1)
@@ -67,17 +68,13 @@ function test_cnn(name, method)
             println("Timing $name...")
             print("cold:")
             @time expl = analyze(input, analyzer)
-
-            @test size(expl.val) == size(input)
             @test_reference "references/cnn/$(name)_max.jld2" Dict("expl" => expl.val) by =
                 (r, a) -> isapprox(r["expl"], a["expl"]; rtol=0.05)
         end
         @testset "Neuron selection" begin
             analyzer = method(model)
             print("warm:")
             @time expl = analyze(input, analyzer, 1)
-
-            @test size(expl.val) == size(input)
             @test_reference "references/cnn/$(name)_ns1.jld2" Dict("expl" => expl.val) by =
                 (r, a) -> isapprox(r["expl"], a["expl"]; rtol=0.05)
         end