New role selectors

[EmilHvitfeldt]

This PR aims to close #942, close #669 and helps with #193.

It does this by expanding the role selectors we have available.

without addressing the *_predictors() variants we used to have

• all_numeric()
• all_nominal()

This PR makes it so we have

• all_numeric()
  • all_integer()
  • all_double()
• all_nominal()
  • all_ordered()
  • all_unordered()
    • all_string()
    • all_factor()
• all_logical()
• all_date()
• all_datetime()

The structure above is to indicate that all_numeric() captures both integers and doubles.

I made the changes to be backwards compatible. The changes are many in a way such that the type variable in var_info is a list column instead of a character column.

TODO:

• Fix tests
• Document that using -all_outcomes() is an anti-pattern
• update steps to work with changes, some steps uses testing like this which doesn't work anymore
• add alll_strings and all_factors

[topepo]

For this to work well, can you take a look at all of the steps that produce indicators (or other integers) and make sure that they all produce actual integers? Otherwise, we'll just beed fielding more issues when this PR is merged.

[EmilHvitfeldt]

The failing build right now is related to some new changes in r-devel related to dates... https://github.com/tidymodels/recipes/actions/runs/3192498090/jobs/5210042479#step:6:234

should be dealt with in a separate PR

[topepo]

Maybe I missed it but can you save a few existing recipes and make sure that it is backwards compatible in unit tests

[EmilHvitfeldt]

This change is backwards compatible in the sense that you an old unprepped recipe is identical to a similarly constructed recipe after being prepped.

# 1.0.1
library(recipes)
rec_spec <- recipe(mpg ~ ., data = mtcars) %>%
  step_poly(disp) %>%
  step_normalize(all_numeric_predictors())

readr::write_rds(rec_spec, "rec_spec.rds")

# This PR
library(recipes)
rec_spec <- readr::read_rds("rec_spec.rds")

new_rec_spec <- recipe(mpg ~ ., data = mtcars) %>%
  step_poly(disp) %>%
  step_normalize(all_numeric_predictors())

identical(prep(rec_spec), prep(new_rec_spec))

You can also bake an old prepped recipe in the new version no problem.

# 1.0.1
library(recipes)
rec_spec <- recipe(mpg ~ ., data = mtcars) %>%
  step_poly(disp) %>%
  step_normalize(all_numeric_predictors()) %>%
  prep()

readr::write_rds(rec_spec, "rec_spec.rds")

# This PR
library(recipes)
rec_spec <- readr::read_rds("rec_spec.rds")

new_rec_spec <- recipe(mpg ~ ., data = mtcars) %>%
  step_poly(disp) %>%
  step_normalize(all_numeric_predictors())  %>%
  prep()

identical(bake(rec_spec, mtcars), bake(new_rec_spec, mtcars))

Where the it brakes down if you try to add steps to a old recipes that has been prepped, and at least one of the steps produced a new variable.

# 1.0.1
library(recipes)
rec_spec <- recipe(mpg ~ ., data = mtcars) %>%
  step_poly(disp) %>%
  step_normalize(all_numeric_predictors()) %>%
  prep()

readr::write_rds(rec_spec, "rec_spec.rds")

# This PR
library(recipes)
rec_spec <- readr::read_rds("rec_spec.rds")

# Doesn't work
rec_spec %>% 
  step_date(all_date()) %>% #assuming that mtcars had a date variable
  prep() %>%
  bake(data = mtcars)

[github-actions]

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.