resample calibration post-processors with an internal split

[simonpcouch]

Related to https://github.com/tidymodels/workflows/pull/225, https://github.com/tidymodels/container/pull/12. This PR proposes resampling calibrators using an "internal split"—it's very scrappy at the moment and intended only for internal testing.

library(tidymodels)
library(container)
library(probably)
#> 
#> Attaching package: 'probably'
#> The following objects are masked from 'package:base':
#> 
#>     as.factor, as.ordered

# create example data
set.seed(1)
dat <- tibble(y = rnorm(100), x = y/2 + rnorm(100))

dat
#> # A tibble: 100 × 2
#>         y      x
#>     <dbl>  <dbl>
#>  1 -0.626 -0.934
#>  2  0.184  0.134
#>  3 -0.836 -1.33 
#>  4  1.60   0.956
#>  5  0.330 -0.490
#>  6 -0.820  1.36 
#>  7  0.487  0.960
#>  8  0.738  1.28 
#>  9  0.576  0.672
#> 10 -0.305  1.53 
#> # ℹ 90 more rows

dat_boots <- bootstraps(dat)

# construct workflow
wf_simple <- workflow(y ~ x, boost_tree("regression", trees = 3))

# specify calibration
reg_ctr <-
  container(mode = "regression") %>%
  adjust_numeric_calibration(type = "linear")

wf_post <- wf_simple %>% add_container(reg_ctr)

# resample workflows
set.seed(1)
wf_simple_res <- 
  fit_resamples(
    wf_simple,
    dat_boots,
    control = control_grid(save_pred = TRUE)
  )

set.seed(1)
wf_post_res <- 
  fit_resamples(
    wf_post,
    dat_boots,
    control = control_grid(save_pred = TRUE)
  )

# ...train the post-processor post-hoc
cal_manual <- cal_estimate_linear(wf_simple_res, truth = y)
cal_manual_preds <- cal_apply(wf_simple_res, cal_manual)

simple_preds <- collect_predictions(wf_simple_res, summarize = TRUE)
cal_auto_preds <- collect_predictions(wf_post_res, summarize = TRUE)
cal_manual_preds
#> # A tibble: 100 × 4
#>      .pred  .row      y .config             
#>      <dbl> <int>  <dbl> <chr>               
#>  1 -0.167      1 -0.626 Preprocessor1_Model1
#>  2  0.267      2  0.184 Preprocessor1_Model1
#>  3  0.215      3 -0.836 Preprocessor1_Model1
#>  4  0.273      4  1.60  Preprocessor1_Model1
#>  5 -0.118      5  0.330 Preprocessor1_Model1
#>  6  0.269      6 -0.820 Preprocessor1_Model1
#>  7  0.140      7  0.487 Preprocessor1_Model1
#>  8  0.219      8  0.738 Preprocessor1_Model1
#>  9  0.254      9  0.576 Preprocessor1_Model1
#> 10  0.0856    10 -0.305 Preprocessor1_Model1
#> # ℹ 90 more rows

Averaged predictions from the uncalibrated model:

ggplot(simple_preds, aes(x = y, y = .pred)) + geom_point()

Averaged predictions from the model calibrated internally in tune:

ggplot(cal_auto_preds, aes(x = y, y = .pred)) + geom_point()

Averaged predictions from the uncalibrated model, calibrated manually after the fact with probably (I’m not sure I got the flow right with cal_estimate_linear(...) %>% cal_apply(...)?):

ggplot(cal_manual_preds, aes(x = y, y = .pred)) + geom_point()

^{Created on 2024-04-26 with reprex v2.1.0}

As-is, this PR doesn't apply any postprocessor if there's not a calibrator in the postprocessor--mostly intended to allow for experimentation on the statistical properties of resampling calibrators in this way.

[topepo]

I was thinking a lot about this this morning. Some thoughts not in our google doc:

• Find a way to control the randomness of the sampling. Before this step, there may be other things that consume random numbers.
• Export all functions involved in this; finetune and others will need to do this (maybe a standalone?)

[simonpcouch]

With an eye for reducing Remotes hoopla, I'm going to go ahead and merge and open issues for smaller todos.

[github-actions[bot]]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.