Small updates

Author: VaclavMacha

docs/src/lecture_11/data/mnist.jl

@@ -2,7 +2,7 @@ using Flux
 using Flux: onecold
 using MLDatasets
 
-include(joinpath(dirname(@__FILE__), "utilities.jl"))
+include(joinpath(dirname(@__FILE__), ("utilities.jl")))
 
 T = Float32
 dataset = MLDatasets.MNIST
@@ -19,7 +19,7 @@ model = Chain(
     softmax,
 )
 
-file_name = joinpath("data", "mnist_sigmoid.jld2")
+file_name = evaldir("mnist_sigmoid.jld2")
 train_or_load!(file_name, model, X_train, y_train)
 
 ii1 = findall(onecold(y_train, 0:9) .== 1)[1:5]

docs/src/lecture_11/data/mnist.jld2

@@ -1,23 +1,33 @@
 using MLDatasets
 using Flux
 
-include(joinpath(dirname(@__FILE__), "utilities.jl"))
+include(joinpath(dirname(@__FILE__), ("utilities.jl")))
 
 dataset = MLDatasets.MNIST
 T = Float32
 X_train, y_train, X_test, y_test = load_data(dataset; T=T, onehot=true)
 
+# model = Chain(
+#     Conv((2, 2), 1 => 16, sigmoid),
+#     MaxPool((2, 2)),
+#     Conv((2, 2), 16 => 8, sigmoid),
+#     MaxPool((2, 2)),
+#     Flux.flatten,
+#     Dense(288, size(y_train, 1)),
+#     softmax,
+# ) |> gpu
+
 model = Chain(
-    Conv((2, 2), 1 => 16, sigmoid),
+    Conv((2, 2), 1 => 16, relu),
     MaxPool((2, 2)),
-    Conv((2, 2), 16 => 8, sigmoid),
+    Conv((2, 2), 16 => 8, relu),
     MaxPool((2, 2)),
     Flux.flatten,
     Dense(288, size(y_train, 1)),
     softmax,
-) |> gpu
+)
 
-file_name = joinpath("data", "mnist_sigmoid.jld2")
-train_model!(model, X_train, y_train; file_name=file_name, n_epochs=100)
+file_name = evaldir("mnist.jld2")
+train_or_load!(file_name, model, X_train, y_train; n_epochs=100, force=true)
 
 accuracy(model, X_test, y_test)

docs/src/lecture_11/data/mnist_gpu.jl

@@ -14,6 +14,8 @@ else
     gpu(x) = x
 end
 
+evaldir(args...) = joinpath(dirname(@__FILE__), args...)
+
 accuracy(model, x, y) = mean(onecold(cpu(model(x))) .== onecold(cpu(y)))
 
 function reshape_data(X::AbstractArray{T,3}, y::AbstractVector) where {T}

docs/src/lecture_11/data/mnist_sigmoid.jld2

@@ -405,7 +405,7 @@ The accuracy is over 93%, which is not bad for training for one epoch only. Let
 println("Test accuracy = " * string(accuracy(X_test, y_test))) # hide
 ```
 
-The externally trained model has an accuracy of more than 98% (it has the same architecture as the one defined above, but it was trained for 50 epochs.). Even though there are perfect models (with accuracy 100%) on MNIST, we are happy with this result. We will perform further analysis of the network in the exercises.
+The externally trained model has an accuracy of almost 98% (it has the same architecture as the one defined above, but it was trained for 100 epochs.). Even though there are perfect models (with accuracy 100%) on MNIST, we are happy with this result. We will perform further analysis of the network in the exercises.
 
 ```@setup nn
 using Plots