Use recipes_map_cols() in bake() methods that transform columns in place

[EmilHvitfeldt]

recipes is gaining an exported recipes_map_cols() helper (tidymodels/recipes#1543, tidymodels/recipes#1556, tidymodels/recipes#1558) for use in bake() methods of steps that transform columns in place. Assigning one column at a time in a for loop (new_data[[col_name]] <- ...) copies the whole data frame on each iteration, which is quadratic in the number of columns. recipes_map_cols() computes all the new columns first and assigns them in bulk. In recipes, baking 8,000 factor columns through step_other() dropped from 5.14s to 1.26s.

The following bake() methods in embed use the slow pattern and could be converted:

• step_lencode() (R/lencode.R:296)
• step_lencode_glm() (R/lencode_glm.R:209)
• step_lencode_bayes() (R/lencode_bayes.R:272)
• step_lencode_mixed() (R/lencode_mixed.R:259)
• step_collapse_stringdist() (R/collapse_stringdist.R:203)
• step_discretize_cart() (R/discretize_cart.R:270)
• step_discretize_xgb() (R/discretize_xgb.R:545)

The two discretize_* steps are the worst of these: they take a full binned_data <- new_data copy and then assign into it once per column, so they pay the copy twice per iteration.

Separately, step_collapse_cart() (R/collapse_cart.R:165) loops per column too, but each iteration calls convert_keys(), which rebuilds the data frame via a join. recipes_map_cols() doesn't drop in directly there, though it is quadratic in its own way and worth a look.

This is blocked on the recipes release that exports recipes_map_cols().