Update Flux dependency to v0.13 (#53)

Author: adrhill

Project.toml

@@ -19,7 +19,7 @@ Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 [compat]
 ColorSchemes = "3"
 Distributions = "0.25"
-Flux = "0.12"
+Flux = "0.12, 0.13"
 ImageCore = "0.8, 0.9"
 PrettyTables = "1"
 Tullio = "0.3"

docs/literate/advanced_lrp.jl

@@ -157,7 +157,7 @@ heatmap(input, analyzer)
 # ### Registering activation functions
 # The mechanism for registering custom activation functions is analogous to that of custom layers:
 myrelu(x) = max.(0, x)
-model = Chain(flatten, Dense(784, 100, myrelu), Dense(100, 10))
+model = Chain(Flux.flatten, Dense(784, 100, myrelu), Dense(100, 10))
 
 # Once again, creating an LRP analyzer for this model will throw an `ArgumentError`
 # and display the following model check summary:

src/flux.jl

@@ -61,11 +61,11 @@ function strip_softmax(model::Chain)
 end
 strip_softmax(l::Union{Dense,Conv}) = set_params(l, l.weight, l.bias, identity)
 
-# helper function to work around Flux.Zeros
+# helper function to work around `bias=false` (Flux v0.13) and `bias=Flux.Zeros` (v0.12)
 function get_params(layer)
     W = layer.weight
     b = layer.bias
-    if typeof(b) <: Flux.Zeros
+    if b == false || typeof(b) <: Flux.Zeros
         b = zeros(eltype(W), size(W, 1))
     end
     return W, b

src/lrp_checks.jl

@@ -78,14 +78,6 @@ function check_model(::Val{:LRP}, c::Chain; verbose=true)
                 LRP_CONFIG.supports_layer(::typeof(mylayer)) = true  # for functions
                 ```
                 The default fallback for this layer will use Automatic Differentiation according to "Layer-Wise Relevance Propagation: An Overview".
-                You can also define a fully LRP-custom rule for your layer by using the interface
-                ```julia
-                function (rule::AbstractLRPRule)(layer::MyLayer, aₖ, Rₖ₊₁)
-                    # ...
-                    return Rₖ
-                end
-                ```
-                This pattern can also be used to dispatch on specific rules.
                 """,
             ),
         )

test/references/rules/EpsilonRule/DepthwiseConv.jld2

@@ -1,4 +1,5 @@
 using Flux
+using Flux: flatten
 using ExplainableAI
 using ExplainableAI: fuse_batchnorm
 using Random

test/references/rules/GammaRule/DepthwiseConv.jld2

@@ -139,7 +139,6 @@ layers = Dict(
     "Conv" => Conv((3, 3), 2 => 4; init=pseudorandn),
     "MaxPool" => MaxPool((3, 3)),
     "MeanPool" => MaxPool((3, 3)),
-    "DepthwiseConv" => DepthwiseConv((3, 3), 2 => 4; init=pseudorandn),
     "ConvTranspose" => ConvTranspose((3, 3), 2 => 4; init=pseudorandn),
     "CrossCor" => CrossCor((3, 3), 2 => 4; init=pseudorandn),
     "flatten" => flatten,