Fit the best model from a tuning grid

complete_tflow(
  x,
  metric,
  ...,
  best_params = NULL,
  method = c("select_best", "select_by_one_std_err", "select_by_pct_loss"),
  control = control_tidyflow()
)

Arguments

x: A tidyflow
metric: The metric of reference from which to pick the best model
...: Extra arguments passed to select_by_one_std_err or select_by_pct_loss
best_params: A 1 row tibble with the best parameters to fit the final model. Should have the same format as the result of select_best, select_by_one_std_err or select_by_pct_loss. If best_params is specified, the method, metric and ... arguments are ignored.
method: which method to use. The possible values are select_best, select_by_one_std_err or select_by_pct_loss. By default, it uses select_best.
control: A control_tidyflow object. The control_parsnip control object inside control_tidyflow is passed to fit.

Value

The tidyflow object updated with the fitted best model. Can be extracted with pull_tflow_fit and used to predict on the training or test data with predict_training or predict_testing

Details

The finalized model is fitted on the training data if plug_split was specified otherwise on the complete data.

Examples

if (FALSE) {
library(parsnip)
library(tune)
library(dials)
library(rsample)

# Fit a regularized regression through a grid search.
reg_mod <- set_engine(linear_reg(penalty = tune(), mixture = tune()),
                      "glmnet")
tuned_res <-
 mtcars %>%
  tidyflow() %>%
  plug_resample(vfold_cv, v = 2) %>%
  plug_formula(mpg ~ .) %>%
  plug_model(reg_mod) %>%
  plug_grid(grid_regular, levels = 1) %>%
  fit()

# Finalize the best model and refit on the whole dataset
final_model <- complete_tflow(tuned_res, metric = "rmse")

# complete_tflow uses tune::select_best as the default method. However,
# tune::select_by_one_std_err and
# tune::select_by_pct_loss can be used. These need to specify the metric and
# the tuning value from which to sort the selection. For example:
final_model_stderr <- complete_tflow(tuned_res,
                                     metric = "rmse",
                                     method = "select_by_one_std_err",
                                     penalty)

# select_by_one_std_err finalizs the best model with the simplest tuning
# values within one standard deviation from most optimal
# combination. For more information on these methods, see
# ?select_best

# You can also specify the best parameters, in case you want
# to override the automatic extraction of the best fit. If you
# specify `best_params` it will override all other arguments

best_params <- select_best(pull_tflow_fit_tuning(tuned_res), metric = "rmse")
final_model_custom <- complete_tflow(tuned_res, best_params = best_params)

# To see the final tuning values, extract the model spec
pull_tflow_spec(final_model)

# To extract the final fitted model:
pull_tflow_fit(final_model)

# Since there was no `plug_split`, the final model is fitted
# entirely on the data (no training/testing). If you try to predict
# on either one, it will not work:
final_model %>%
  predict_training()

# Add a split step, fit again and then finalize the model
# to predict on the training set
tuned_split <-
  tuned_res %>%
  replace_grid(grid_regular) %>%
  plug_split(initial_split) %>%
  fit()

tuned_split %>%
 complete_tflow(metric = "rmse") %>%
 predict_training()
}