Support Parallel layers in LRP

[adrhill]

Closes #10, building on top of infrastructure introduced in #119. 🎉

Examples

Using Metalhead's pre-trained ResNet(152) model without canonization on the Readme example.
Input for reference:

Heapmaps for class "castle"

Using EpsilonPlusFlat composite:

Using EpsilonPlus composite:

Heapmaps for class "street sign"

Using EpsilonPlusFlat composite:

Using EpsilonPlus composite:

Code used to generate heatmaps:

using ExplainableAI
using Flux
using Metalhead                         # pre-trained vision models
using HTTP, FileIO, ImageMagick         # load image from URL
using ImageInTerminal

# Load input
url = HTTP.URI("https://raw.githubusercontent.com/adrhill/ExplainableAI.jl/gh-pages/assets/heatmaps/castle.jpg")
img = load(url)
input = preprocess_imagenet(img)
input = reshape(input, 224, 224, 3, :)  # reshape to WHCN format

# Define model sizes
resnet(size) = ResNet(size, pretrain=true).layers 
resnet_sizes = (18, 50, 152)

# Load composites
composites = Dict(
    "EpsilonPlus"            => EpsilonPlus(),  
    "EpsilonPlusFlat"        => EpsilonPlusFlat(), 
    "EpsilonAlpha2Beta1"     => EpsilonAlpha2Beta1(), 
    "EpsilonAlpha2Beta1Flat" => EpsilonAlpha2Beta1Flat(),
)


# Generate heatmaps for classes castle and street sign
for s in resnet_sizes
    model = resnet(s)
    for (cname, c) in composites
        @info "Analyzing ResNet$s on composite $cname" 
        analyzer = LRP(model, c)
        h_castle = heatmap(input, analyzer)
        h_stsign = heatmap(input, analyzer, 920)
        
        display(h_castle)
        display(h_stsign)
        save("resnet$(s)_$(cname)_castle.png", h_castle)
        save("resnet$(s)_$(cname)_stsign.png", h_stsign)
    end
end