finetune

R-CMD-check Codecov test coverage Lifecycle

finetune contains some extra functions for model tuning that extend what is currently in the tune package.

Very rough version of the package right now but it works fairly well. There are two main sets of tools.

Tuning via simulated annealing optimization is another iterative search tool for finding good values:

library(tidymodels)
library(finetune)

# Syntax very similar to `tune_grid()` or `tune_Bayes()`: 

## -----------------------------------------------------------------------------

data(two_class_dat, package = "modeldata")

set.seed(1)
rs <- bootstraps(two_class_dat, times = 10) # more resamples usually needed

# Optimize a regularized discriminant analysis model
library(discrim)
rda_spec <-
  discrim_regularized(frac_common_cov = tune(), frac_identity = tune()) %>%
  set_engine("klaR")

## -----------------------------------------------------------------------------

ctrl <- control_sim_anneal(verbose = TRUE)

set.seed(2)
sa_res <- 
  rda_spec %>% 
  tune_sim_anneal(Class ~ ., resamples = rs, iter = 20, initial = 4, control = ctrl)
#> Optimizing roc_auc
#> Initial best: 0.86480
#>  1 ♥ new best           roc_auc=0.87327  (+/-0.004592)
#>  2 ♥ new best           roc_auc=0.87915  (+/-0.003864)
#>  3 ◯ accept suboptimal  roc_auc=0.87029  (+/-0.004994)
#>  4 + better suboptimal  roc_auc=0.87171  (+/-0.004717)
#>  5 ◯ accept suboptimal  roc_auc=0.86944  (+/-0.005081)
#>  6 ◯ accept suboptimal  roc_auc=0.86812  (+/-0.0052)
#>  7 ♥ new best           roc_auc=0.88172  (+/-0.003647)
#>  8 ◯ accept suboptimal  roc_auc=0.87678  (+/-0.004276)
#>  9 ◯ accept suboptimal  roc_auc=0.8627   (+/-0.005784)
#> 10 + better suboptimal  roc_auc=0.87003  (+/-0.005106)
#> 11 + better suboptimal  roc_auc=0.87088  (+/-0.004962)
#> 12 ◯ accept suboptimal  roc_auc=0.86803  (+/-0.005195)
#> 13 ◯ accept suboptimal  roc_auc=0.85294  (+/-0.006498)
#> 14 ─ discard suboptimal roc_auc=0.84689  (+/-0.006867)
#> 15 ✖ restart from best  roc_auc=0.85021  (+/-0.006623)
#> 16 ◯ accept suboptimal  roc_auc=0.87607  (+/-0.004318)
#> 17 ◯ accept suboptimal  roc_auc=0.87245  (+/-0.004799)
#> 18 + better suboptimal  roc_auc=0.87706  (+/-0.004131)
#> 19 ◯ accept suboptimal  roc_auc=0.87213  (+/-0.004791)
#> 20 ◯ accept suboptimal  roc_auc=0.86218  (+/-0.005773)
show_best(sa_res, metric = "roc_auc", n = 2)
#> # A tibble: 2 × 9
#>   frac_common_cov frac_ident…¹ .metric .esti…²  mean     n std_err .config .iter
#>             <dbl>        <dbl> <chr>   <chr>   <dbl> <int>   <dbl> <chr>   <int>
#> 1           0.308       0.0166 roc_auc binary  0.882    10 0.00365 Iter7       7
#> 2           0.121       0.0474 roc_auc binary  0.879    10 0.00386 Iter2       2
#> # … with abbreviated variable names ¹​frac_identity, ²​.estimator

The second set of methods are for “racing”. We start off by doing a small set of resamples for all of the grid points, then statistically testing to see which ones should be dropped or investigated more. The two methods here are based on those should in Kuhn (2014).

For example, using an ANOVA-type analysis to filter out parameter combinations:

set.seed(3)
grid <-
  rda_spec %>%
  extract_parameter_set_dials() %>%
  grid_max_entropy(size = 20)

ctrl <- control_race(verbose_elim = TRUE)

set.seed(4)
grid_anova <- 
  rda_spec %>% 
  tune_race_anova(Class ~ ., resamples = rs, grid = grid, control = ctrl)
#> ℹ Racing will maximize the roc_auc metric.
#> ℹ Resamples are analyzed in a random order.
#> ℹ Bootstrap10: 14 eliminated;  6 candidates remain.
#> ℹ Bootstrap04:  2 eliminated;  4 candidates remain.
#> ℹ Bootstrap03: All but one parameter combination were eliminated.

show_best(grid_anova, metric = "roc_auc", n = 2)
#> # A tibble: 2 × 9
#>   .order frac_common_cov frac_iden…¹ .metric .esti…²  mean     n std_err .config
#>    <int>           <dbl>       <dbl> <chr>   <chr>   <dbl> <int>   <dbl> <chr>  
#> 1      9           0.831      0.0207 roc_auc binary  0.901     1      NA Prepro…
#> 2      8           0.831      0.0207 roc_auc binary  0.893     1      NA Prepro…
#> # … with abbreviated variable names ¹​frac_identity, ²​.estimator

tune_race_win_loss() can also be used. It treats the tuning parameters as sports teams in a tournament and computed win/loss statistics.

set.seed(4)
grid_win_loss<- 
  rda_spec %>% 
  tune_race_win_loss(Class ~ ., resamples = rs, grid = grid, control = ctrl)
#> ℹ Racing will maximize the roc_auc metric.
#> ℹ Resamples are analyzed in a random order.
#> ℹ Bootstrap10:  3 eliminated; 17 candidates remain.
#> ℹ Bootstrap04:  2 eliminated; 15 candidates remain.
#> ℹ Bootstrap03:  2 eliminated; 13 candidates remain.
#> ℹ Bootstrap01:  1 eliminated; 12 candidates remain.
#> ℹ Bootstrap07:  1 eliminated; 11 candidates remain.
#> ℹ Bootstrap05:  1 eliminated; 10 candidates remain.
#> ℹ Bootstrap08:  1 eliminated;  9 candidates remain.

show_best(grid_win_loss, metric = "roc_auc", n = 2)
#> # A tibble: 2 × 9
#>   .order frac_common_cov frac_iden…¹ .metric .esti…²  mean     n std_err .config
#>    <int>           <dbl>       <dbl> <chr>   <chr>   <dbl> <int>   <dbl> <chr>  
#> 1      9           0.831      0.0207 roc_auc binary  0.901     1      NA Prepro…
#> 2      9           0.321      0.0561 roc_auc binary  0.898     1      NA Prepro…
#> # … with abbreviated variable names ¹​frac_identity, ²​.estimator

Contributing

This project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.