Use DifferentiationInterface for gradient-based analyzers (#167)

Author: adrhill

CHANGELOG.md

@@ -1,4 +1,9 @@
 # ExplainableAI.jl
+## Version `v0.9.0`
+- ![Feature][badge-feature] Support selection of AD backend via DifferentiationInterface.jl ([#167])
+  - `Gradient`, `InputTimesGradient` and `GradCAM` analyzers now have an additional `backend` field and type parameter
+- ![Maintenance][badge-maintenance] Update XAIBase interface to v4 ([#166])
+
 ## Version `v0.8.0`
 This release removes the automatic reexport of heatmapping functionality.
 Users are now required to manually load 
@@ -210,6 +215,8 @@ Performance improvements:
 [VisionHeatmaps]: https://julia-xai.github.io/XAIDocs/VisionHeatmaps/stable/
 [TextHeatmaps]: https://julia-xai.github.io/XAIDocs/TextHeatmaps/stable/
 
+[#167]: https://github.com/Julia-XAI/ExplainableAI.jl/pull/167
+[#166]: https://github.com/Julia-XAI/ExplainableAI.jl/pull/166
 [#162]: https://github.com/Julia-XAI/ExplainableAI.jl/pull/162
 [#159]: https://github.com/Julia-XAI/ExplainableAI.jl/pull/159
 [#157]: https://github.com/Julia-XAI/ExplainableAI.jl/pull/157

Project.toml

@@ -1,9 +1,11 @@
 name = "ExplainableAI"
 uuid = "4f1bc3e1-d60d-4ed0-9367-9bdff9846d3b"
 authors = ["Adrian Hill <gh@adrianhill.de>"]
-version = "0.8.1"
+version = "0.9.0-DEV"
 
 [deps]
+ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"
+DifferentiationInterface = "a0c0ee7d-e4b9-4e03-894e-1c5f64a51d63"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
@@ -12,6 +14,8 @@ XAIBase = "9b48221d-a747-4c1b-9860-46a1d8ba24a7"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [compat]
+ADTypes = "1"
+DifferentiationInterface = "0.5"
 Distributions = "0.25"
 Random = "<0.0.1, 1"
 Reexport = "1"

src/ExplainableAI.jl

@@ -7,7 +7,12 @@ import XAIBase: call_analyzer
 using Base.Iterators
 using Distributions: Distribution, Sampleable, Normal
 using Random: AbstractRNG, GLOBAL_RNG
+
+# Automatic differentiation
+using ADTypes: AbstractADType, AutoZygote
+using DifferentiationInterface: value_and_pullback
 using Zygote
+const DEFAULT_AD_BACKEND = AutoZygote()
 
 include("compat.jl")
 include("bibliography.jl")

src/gradcam.jl

@@ -15,16 +15,25 @@ GradCAM is compatible with a wide variety of CNN model-families.
 # References
 - $REF_SELVARAJU_GRADCAM
 """
-struct GradCAM{F,A} <: AbstractXAIMethod
+struct GradCAM{F,A,B<:AbstractADType} <: AbstractXAIMethod
     feature_layers::F
     adaptation_layers::A
+    backend::B
+
+    function GradCAM(
+        feature_layers::F, adaptation_layers::A, backend::B=DEFAULT_AD_BACKEND
+    ) where {F,A,B<:AbstractADType}
+        new{F,A,B}(feature_layers, adaptation_layers, backend)
+    end
 end
 function call_analyzer(input, analyzer::GradCAM, ns::AbstractOutputSelector; kwargs...)
     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)
+    grad, output, output_indices = gradient_wrt_input(
+        analyzer.adaptation_layers, A, ns, analyzer.backend
+    )
     αᶜ = sum(grad; dims=(1, 2)) / feature_map_size
     Lᶜ = max.(sum(αᶜ .* A; dims=3), 0)
     return Explanation(Lᶜ, input, output, output_indices, :GradCAM, :cam, nothing)

src/gradient.jl

@@ -1,26 +1,45 @@
-function gradient_wrt_input(model, input, ns::AbstractOutputSelector)
-    output, back = Zygote.pullback(model, input)
-    output_indices = ns(output)
-
-    # Compute VJP w.r.t. full model output, selecting vector s.t. it masks output neurons
-    v = zero(output)
-    v[output_indices] .= 1
-    grad = only(back(v))
-    return grad, output, output_indices
+function forward_with_output_selection(model, input, selector::AbstractOutputSelector)
+    output = model(input)
+    sel = selector(output)
+    return output[sel]
+end
+
+function gradient_wrt_input(
+    model, input, output_selector::AbstractOutputSelector, backend::AbstractADType
+)
+    output = model(input)
+    return gradient_wrt_input(model, input, output, output_selector, backend)
+end
+
+function gradient_wrt_input(
+    model, input, output, output_selector::AbstractOutputSelector, backend::AbstractADType
+)
+    output_selection = output_selector(output)
+    dy = zero(output)
+    dy[output_selection] .= 1
+
+    output, grad = value_and_pullback(model, backend, input, dy)
+    return grad, output, output_selection
 end
 
 """
     Gradient(model)
 
 Analyze model by calculating the gradient of a neuron activation with respect to the input.
 """
-struct Gradient{M} <: AbstractXAIMethod
+struct Gradient{M,B<:AbstractADType} <: AbstractXAIMethod
     model::M
-    Gradient(model) = new{typeof(model)}(model)
+    backend::B
+
+    function Gradient(model::M, backend::B=DEFAULT_AD_BACKEND) where {M,B<:AbstractADType}
+        new{M,B}(model, backend)
+    end
 end
 
 function call_analyzer(input, analyzer::Gradient, ns::AbstractOutputSelector; kwargs...)
-    grad, output, output_indices = gradient_wrt_input(analyzer.model, input, ns)
+    grad, output, output_indices = gradient_wrt_input(
+        analyzer.model, input, ns, analyzer.backend
+    )
     return Explanation(
         grad, input, output, output_indices, :Gradient, :sensitivity, nothing
     )
@@ -32,15 +51,23 @@ end
 Analyze model by calculating the gradient of a neuron activation with respect to the input.
 This gradient is then multiplied element-wise with the input.
 """
-struct InputTimesGradient{M} <: AbstractXAIMethod
+struct InputTimesGradient{M,B<:AbstractADType} <: AbstractXAIMethod
     model::M
-    InputTimesGradient(model) = new{typeof(model)}(model)
+    backend::B
+
+    function InputTimesGradient(
+        model::M, backend::B=DEFAULT_AD_BACKEND
+    ) where {M,B<:AbstractADType}
+        new{M,B}(model, backend)
+    end
 end
 
 function call_analyzer(
     input, analyzer::InputTimesGradient, ns::AbstractOutputSelector; kwargs...
 )
-    grad, output, output_indices = gradient_wrt_input(analyzer.model, input, ns)
+    grad, output, output_indices = gradient_wrt_input(
+        analyzer.model, input, ns, analyzer.backend
+    )
     attr = input .* grad
     return Explanation(
         attr, input, output, output_indices, :InputTimesGradient, :attribution, nothing