View Source Scholar.ModelSelection (Scholar v0.3.1)

Module containing cross validation, splitting function, and other model selection methods.

Summary

Functions

General interface of cross validation.

General interface of grid search.

Perform K-Fold split on the given data.

General interface of weighted cross validation.

General interface of weighted grid search.

Functions

Link to this function

cross_validate(x, y, folding_fun, scoring_fun)

View Source

General interface of cross validation.

Examples

iex> folding_fun = fn x -> Scholar.ModelSelection.k_fold_split(x, 3) end
iex> scoring_fun = fn x, y ->
...>   {x_train, x_test} = x
...>   {y_train, y_test} = y
...>   model = Scholar.Linear.LinearRegression.fit(x_train, y_train, fit_intercept?: true)
...>   y_pred = Scholar.Linear.LinearRegression.predict(model, x_test)
...>   mse = Scholar.Metrics.Regression.mean_square_error(y_test, y_pred)
...>   mae = Scholar.Metrics.Regression.mean_absolute_error(y_test, y_pred)
...>   [mse, mae]
...> end
iex> x = Nx.iota({7, 2})
iex> y = Nx.tensor([0, 1, 2, 0, 1, 1, 0])
iex> Scholar.ModelSelection.cross_validate(x, y, folding_fun, scoring_fun)
#Nx.Tensor<
  f32[2][3]
  [
    [1.5700000524520874, 1.2149654626846313, 0.005000002216547728],
    [1.100000023841858, 1.0735294818878174, 0.050000011920928955]
  ]
>
Link to this function

grid_search(x, y, folding_fun, scoring_fun, opts)

View Source

General interface of grid search.

The opts must be a keyword list of list values, which will become different combinations to perform the grid search on.

Examples

iex> folding_fun = fn x -> Scholar.ModelSelection.k_fold_split(x, 3) end
iex> scoring_fun = fn x, y, opts ->
...>   {x_train, x_test} = x
...>   {y_train, y_test} = y
...>   model = Scholar.Linear.LogisticRegression.fit(x_train, y_train, opts)
...>   y_pred = Scholar.Linear.LogisticRegression.predict(model, x_test)
...>   mse = Scholar.Metrics.Regression.mean_square_error(y_test, y_pred)
...>   mae = Scholar.Metrics.Regression.mean_absolute_error(y_test, y_pred)
...>   [mse, mae]
...> end
iex> x = Nx.iota({7, 2})
iex> y = Nx.tensor([0, 1, 2, 0, 1, 1, 0])
iex> opts = [
...>   num_classes: [3],
...>   iterations: [10, 20, 50],
...>   optimizer: [Polaris.Optimizers.adam(learning_rate: 0.005), Polaris.Optimizers.adam(learning_rate: 0.01)],
...> ]
iex> Scholar.ModelSelection.grid_search(x, y, folding_fun, scoring_fun, opts)

Perform K-Fold split on the given data.

Examples

iex> x = Nx.iota({7, 2})
iex> Scholar.ModelSelection.k_fold_split(x, 2) |> Enum.to_list()
[
  {Nx.tensor(
    [
      [6, 7],
      [8, 9],
      [10, 11]
    ]
  ),
  Nx.tensor(
    [
      [0, 1],
      [2, 3],
      [4, 5]
    ]
  )},
  {Nx.tensor(
    [
      [0, 1],
      [2, 3],
      [4, 5]
    ]
  ),
  Nx.tensor(
    [
      [6, 7],
      [8, 9],
      [10, 11]
    ]
  )}
]
Link to this function

weighted_cross_validate(x, y, weights, folding_fun, scoring_fun)

View Source

General interface of weighted cross validation.

Examples

iex> folding_fun = fn x -> Scholar.ModelSelection.k_fold_split(x, 3) end
iex> scoring_fun = fn x, y, weights ->
...>   {x_train, x_test} = x
...>   {y_train, y_test} = y
...>   {weights_train, _weights_test} = weights
...>   model = Scholar.Linear.LinearRegression.fit(x_train, y_train, fit_intercept?: true, sample_weights: weights_train)
...>   y_pred = Scholar.Linear.LinearRegression.predict(model, x_test)
...>   mse = Scholar.Metrics.Regression.mean_square_error(y_test, y_pred)
...>   mae = Scholar.Metrics.Regression.mean_absolute_error(y_test, y_pred)
...>   [mse, mae]
...> end
iex> x = Nx.iota({7, 2})
iex> y = Nx.tensor([0, 1, 2, 0, 1, 1, 0])
iex> weights = Nx.tensor([1, 2, 1, 2, 1, 2, 1])
iex> Scholar.ModelSelection.weighted_cross_validate(x, y, weights, folding_fun, scoring_fun)
#Nx.Tensor<
  f32[2][3]
  [
    [0.5010337233543396, 1.1419668197631836, 0.35123950242996216],
    [0.522727370262146, 1.0526316165924072, 0.590908944606781]
  ]
>
Link to this function

weighted_grid_search(x, y, weights, folding_fun, scoring_fun, opts)

View Source

General interface of weighted grid search.

If you want to use opts in some functions inside scoring_fun, you need to pass it as a parameter like in the example below.

Examples

iex> folding_fun = fn x -> Scholar.ModelSelection.k_fold_split(x, 3) end
iex> scoring_fun = fn x, y, weights, opts ->
...>   {x_train, x_test} = x
...>   {y_train, y_test} = y
...>   {weights_train, _weights_test} = weights
...>   opts = Keyword.put(opts, :sample_weights, weights_train)
...>   model = Scholar.Linear.RidgeRegression.fit(x_train, y_train, opts)
...>   y_pred = Scholar.Linear.RidgeRegression.predict(model, x_test)
...>   mse = Scholar.Metrics.Regression.mean_square_error(y_test, y_pred)
...>   mae = Scholar.Metrics.Regression.mean_absolute_error(y_test, y_pred)
...>   [mse, mae]
...> end
iex> x = Nx.iota({7, 2})
iex> y = Nx.tensor([0, 1, 2, 0, 1, 1, 0])
iex> weights = [Nx.tensor([1, 2, 1, 2, 1, 2, 1]), Nx.tensor([2, 1, 2, 1, 2, 1, 2])]
iex> opts = [
...>   alpha: [0, 1, 5],
...>   fit_intercept?: [true, false],
...> ]
iex> Scholar.ModelSelection.weighted_grid_search(x, y, weights, folding_fun, scoring_fun, opts)