Merge branch 'main' of https://github.com/JohannesNaegele/MaNN.jl

.github/workflows/CI.yml

@@ -18,7 +18,7 @@ jobs:
       fail-fast: false
       matrix:
         version:
-          - '1.0'
+          - '1.6'
           - '1.9'
           - 'nightly'
         os:

Project.toml

@@ -9,7 +9,10 @@ MLDatasets = "eb30cadb-4394-5ae3-aed4-317e484a6458"
 Plots = "91a5bcdd-55d7-5caf-9e0b-520d859cae80"
 
 [compat]
-julia = "1"
+Plots = "1"
+MLDatasets = "0.7"
+BenchmarkTools = "1"
+julia = "1.6"
 
 [extras]
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

src/Chain.jl

@@ -19,5 +19,5 @@ This function return the parameters of our layers.
 However, it might make sense to change this implementation with the introduction of autodiff.
 """
 function params(chain::Chain{T}) where T
-    return [params(layer) for layer in chain.layers]
+    return vcat([params(layer) for layer in chain.layers]...)
 end

src/Metrics.jl

@@ -1,4 +1,5 @@
 abstract type AbstractHook end
 abstract type AbstractStage end
 
-# TODO: define hooks and stages
+# TODO: define hooks and stages
+# example commit

src/Optimizer.jl

@@ -7,8 +7,6 @@ The type/layout of params might get changed in the future.
 """
 function update!(opt::BoringOptimizer, params, grad)
     for i in eachindex(params)
-        for j in eachindex(params[i])
-            params[i][j] .-= opt.learning_rate .* grad[i][j]
-        end
+        params[i] .-= opt.learning_rate .* grad[i]
     end
 end

src/Train.jl

@@ -13,7 +13,7 @@ function gradient_hardcoded(chain::Chain, loss, label, input::Vector{Vector{T}},
         layer = chain.layers[i]
         if layer isa AbstractLayer
             δ .= δ .* layer.activation.(input[i + 1], derivative=true)
-            gs[i] = [δ * input[i]', δ]
+            gs[2 * i - 1], gs[2 * i] = δ * input[i]', δ
             δ = layer.weights' * δ
         else # if it is softmax: do nothing
             error_msg = "This backpropagation method works only for a specific configuration."