SHAP Example

The SHAP call is very similar to the LIME call. For the example, we load an image from the imagenet-sample-images repository. An explanation will be generated for the image later on.

using ExplainableAI
using Metalhead: ResNet
using JML_XAI_Project
using Images
using VisionHeatmaps

img = load("images/dog.jpeg")

Image pre-processing

The image is processed in order to use it as input for a model.

imgVec = permutedims(channelview(img),(3,2,1));
imgVec = reshape(imgVec, size(imgVec)..., 1);
input = Float32.(imgVec);

Generation of the explanation

The next step is to initialize a pre-trained ResNet model and apply SHAP to it.

model = ResNet(18; pretrain = true);
model = model.layers;

#A different kernel function is used to call SHAP,
#and no feature selection is required, which is why LASSO must be set to false
analyzer = LIME(model, agnostic_kernel, false)
expl = analyze(input, analyzer);

XAIBase.Explanation{Array{Float64, 4}, Matrix{Float32}, CartesianIndex{2}, Nothing}([0.0 0.0 … 0.0 0.0; 0.0 0.0 … 0.0 0.0; … ; 0.0 0.0 … 0.0 0.0; 0.0 0.0 … 0.0 0.0;;;;], Float32[1.0904795; -0.4928558; … ; 1.3466196; 1.1687055;;], CartesianIndex(239, 1), :MyMethod, :attribution, nothing)

Visualize explanaition

The generated explanation can now be displayed as a heat map.

using VisionHeatmaps
heatmap(expl.val)

function generate_heatmap(map; img=nothing, overlay=false, blurring=false, gaussSTD=2)
    map = heatmap(map.val)

    if blurring == true
        gaussKern2 = ImageFiltering.KernelFactors.gaussian((gaussSTD,gaussSTD))
        map = ImageFiltering.imfilter(map, gaussKern2)
    end

    if overlay == true
        map = (0.5.*Gray.(img) + 0.5.*map)
    end

    return map
end

generate_heatmap (generic function with 1 method)

generate_heatmap(expl, img=img, overlay=true, blurring=true)

Label

To find out the generated corresponding label of the image, we output the number of the label, which can be looked up in the linked text file.

print("Label: ", argmax(expl.output[:,1]) - 1)

Label: 238