Chapter One

Mix.install([
  {:axon, "~> 0.6"},
  {:nx, "~> 0.6"},
  {:explorer, "~> 0.7"},
  {:kino, "~> 0.10"}
])

Get The Iris Dataset

require Explorer.DataFrame, as: DF

iris = Explorer.Datasets.iris()

Extract Train and Test Data

shuffled_normalized_iris =
  iris
  |> DF.mutate(
    for col <- across(~w(sepal_length sepal_width petal_length petal_width)) do
      {col.name, col - mean(col) / variance(col)}
    end
  )
  |> DF.shuffle()

train_iris = DF.slice(shuffled_normalized_iris, 0..119)
test_iris = DF.slice(shuffled_normalized_iris, 120..149)

Extract Train and Test Features and Labels

feature_columns = ~w(sepal_length sepal_width petal_length petal_width)

train_x = Nx.stack(train_iris[feature_columns], axis: 1)

train_y =
  train_iris["species"]
  |> Explorer.Series.cast(:category)
  |> Nx.stack(axis: -1)
  |> Nx.equal(Nx.iota({1, 3}, axis: -1))

test_x = Nx.stack(test_iris[feature_columns], axis: 1)

test_y =
  test_iris["species"]
  |> Explorer.Series.cast(:category)
  |> Nx.stack(axis: -1)
  |> Nx.equal(Nx.iota({1, 3}, axis: -1))

Declare the Model

model =
  Axon.input("iris_features")
  |> Axon.dense(3, activation: :softmax)

Axon.Display.as_graph(model, Nx.template({1, 4}, :f32))

Train the Model

data_stream = Stream.repeatedly(fn -> {train_x, train_y} end)

trained_model_state =
  model
  |> Axon.Loop.trainer(:categorical_cross_entropy, :sgd)
  |> Axon.Loop.metric(:accuracy)
  |> Axon.Loop.run(data_stream, %{}, iterations: 500, epochs: 10)

Evaluate the Model

data = [{test_x, test_y}]

evaluation =
  model
  |> Axon.Loop.evaluator()
  |> Axon.Loop.metric(:accuracy)
  |> Axon.Loop.run(data, trained_model_state)