In awconway/eicu_nitro: What the Package Does (One Line, Title Case)

library(targets)
library(tidymodels)

Recipe for the model

```{targets recBoost} tar_target(recBoost, recipe(sbp_post ~ sbp_pre + nitro_pre + nitro_diff_mcg + total_nitro + n_nitro + nitro_time + time_since_nitro + sbp_pre_mean_60 + sbp_pre_sd_60 + pain_score, data = training )|> step_dummy(all_nominal_predictors()) )

## Model specification

```{targets specBoost}
tar_target(
    specBoost,
    boost_tree(
      # trees = 500,
      # min_n = 30,
      # learn_rate = 0.015,
      # sample_size = 0.5, #0.5 was best
      # tree_depth = 1,
      trees = tune(),
      min_n = tune(),
      mtry = tune(),
      learn_rate = tune(),
      sample_size = tune(),
      tree_depth = tune(),
      loss_reduction = tune()
    ) |>
      set_engine("xgboost")|>
      set_mode("regression")
  )

Model workflow

```{targets workflowBoost} tar_target( workflowBoost, workflow()|> add_recipe(recBoost)|> add_model(specBoost) )

Grid search ranges for tuning model parameters

```{targets gridBoost}
tar_target(gridBoost,
  grid_max_entropy(
    tree_depth(c(8, 15)),
    min_n(c(20L, 40L)),
    mtry(c(4L, 6L)),
    loss_reduction(),
    sample_size = sample_prop(c(0.5,1.0)),
    learn_rate(c(-3,-2)),
    trees(c(500, 5000)),
    size = 50
    )
  )

The tune_race_anova function can be used to speed up the process of finding suitable parameters for the model. Results are returned from 5-fold cross validation.

```{targets tuningRaceBoost} tar_target(tuningRaceBoost, tune_race_anova( workflowBoost, foldsFive, grid = gridBoost, metrics = mset, control = control_race(verbose_elim = TRUE) ) )

A simple tuning grid is useful when interested in tuning just a couple of parameters.

```{targets tuningGridBoost}
# tar_target(tuningGridBoost,
#       tune_grid(
#        workflowBoost,
#         foldsFive,
#         grid = crossing(
#           min_n = seq(30,40, 5)
#     ),
#         metrics = mset,
#         control = control
#       )
#   )

Once parameters are chosen, we can check how they performed with 5-fold cross-validation.

```{targets resampleBoost}

tar_target(resampleBoost,

fit_resamples(

specBoost,

recBoost,

resamples = foldsFive,

control = controlResamples

)

)

## Final model evaluation

The best model from cross-validation is chosen to update the final workflow. The last_fit function trains and then tests with the testing data. 

**Only run when predictor set and parameters for the model are selected.**

```{targets boostFinal, tar_simple = TRUE}
# best <-  select_by_one_std_err(tuningRaceBoost, "learn_rate")
# boost_best <- finalize_workflow(
#   workflowBoost, best
# )
# last_fit(boost_best, initSplit)

awconway/eicu_nitro documentation built on Feb. 21, 2022, 4:35 p.m.