Add ptype information to recipes

[EmilHvitfeldt]

This looks at how we could add ptype information to the recipe.

I think the best plan of action is to take a ptype slice of the entire dataset passed to recipe() and then have helper functions to pull out the needed information.

Because the ptype information that the user/functions need will depend:

• ptype checking for the data passed to prep() should match the data passed to recipe()
• ptype checking for for data passed to bake() shouldn't match against outcome roles
• In essence, everything will depend on the roles, and work with update_role_requirements()

Revdepcheck

I ran revdepcheck on the change to add a ptype field to the recipe object and it came back with no issues

[EmilHvitfeldt]

@juliasilge is this enough for what you need in {vetiver}?

[simonpcouch]

In the original, seeing from Davis:

This isn't a huge deal in practice, because workflows/hardhat typically handle the type consistency ahead of time

And then from Emil:

slightly bigger deal because {workflows}/{hardhat} doesn't deal with this

A bit confused here. From what I can tell in the reprex in #793 (comment), workflows does error in this situation but does so uninformatively. That same uninformative error still occurs with this recipes PR installed—here's the same linked reprex with this PR:

library(recipes)
#> Loading required package: dplyr
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
#> 
#> Attaching package: 'recipes'
#> The following object is masked from 'package:stats':
#> 
#>     step

traning <- tibble(outcome = rnorm(1000), x = rnorm(1000), y = sample(letters, 1000, T))

rec1 <- recipe(outcome ~ ., traning) %>%
  step_bin2factor(where(is.numeric), - all_outcomes())

# Original template `traning` has a character `y`
rec2 <- recipe(outcome ~ ., traning) %>%
  step_bin2factor(all_numeric_predictors())

# But when we prep we have a numeric `y`
testing <- tibble(outcome = rnorm(1000), x = rnorm(1000), y = rnorm(1000))

prep(rec1, training = testing, fresh = TRUE)
#> 
#> ── Recipe ──────────────────────────────────────────────────────────────────────
#> 
#> ── Inputs
#> Number of variables by role
#> outcome:   1
#> predictor: 2
#> 
#> ── Training information
#> Training data contained 1000 data points and no incomplete rows.
#> 
#> ── Operations
#> • Dummy variable to factor conversion for: x and y | Trained

prep(rec2, training = testing, fresh = TRUE)
#> 
#> ── Recipe ──────────────────────────────────────────────────────────────────────
#> 
#> ── Inputs
#> Number of variables by role
#> outcome:   1
#> predictor: 2
#> 
#> ── Training information
#> Training data contained 1000 data points and no incomplete rows.
#> 
#> ── Operations
#> • Dummy variable to factor conversion for: x | Trained

prep(rec1, training = testing, fresh = FALSE)
#> 
#> ── Recipe ──────────────────────────────────────────────────────────────────────
#> 
#> ── Inputs
#> Number of variables by role
#> outcome:   1
#> predictor: 2
#> 
#> ── Training information
#> Training data contained 1000 data points and no incomplete rows.
#> 
#> ── Operations
#> • Dummy variable to factor conversion for: x and y | Trained

prep(rec2, training = testing, fresh = FALSE)
#> 
#> ── Recipe ──────────────────────────────────────────────────────────────────────
#> 
#> ── Inputs
#> Number of variables by role
#> outcome:   1
#> predictor: 2
#> 
#> ── Training information
#> Training data contained 1000 data points and no incomplete rows.
#> 
#> ── Operations
#> • Dummy variable to factor conversion for: x | Trained

rec1$var_info$type
#> [[1]]
#> [1] "double"  "numeric"
#> 
#> [[2]]
#> [1] "string"    "unordered" "nominal"  
#> 
#> [[3]]
#> [1] "double"  "numeric"

rec2$var_info$type
#> [[1]]
#> [1] "double"  "numeric"
#> 
#> [[2]]
#> [1] "string"    "unordered" "nominal"  
#> 
#> [[3]]
#> [1] "double"  "numeric"

library(tidymodels)

wf1 <- workflow(rec1, linear_reg())
wf2 <- workflow(rec2, linear_reg())

fit(wf1, testing)
#> Error in `contrasts<-`(`*tmp*`, value = contr.funs[1 + isOF[nn]]): contrasts can be applied only to factors with 2 or more levels

fit(wf2, testing)
#> Error in `contrasts<-`(`*tmp*`, value = contr.funs[1 + isOF[nn]]): contrasts can be applied only to factors with 2 or more levels

^{Created on 2024-06-03 with reprex v2.1.0}

[simonpcouch]

Could you add some context here on how you imagine this helper being integrated into recipes/workflows?

[EmilHvitfeldt]

Could you add some context here on how you imagine this helper being integrated into recipes/workflows?

right now we have check_training_set() that is called in recipes to check the incoming data, but it only checks column names. recipes_ptype() added in this PR is planned to be used in check_training_set() to enhance checking of input to include types and attributes.

in short, this PR doesn't close 793, but adds the tools needed for it to be closed.

[topepo]

Looks great and it will help with some S3 generics that use workflows (it's hard to know what the original data was in a workflow).

I wanted to verify that we computed the ptype after the formula was used to subset the columns in the recipe 👍

[EmilHvitfeldt]

I wanted to verify that we computed the ptype after the formula was used to subset the columns in the recipe 👍

That is happening, but I see I forgot to add a test to verify.

[juliasilge]

In vetiver, we currently require users to pass in their "input prototype data" explicitly, as shown in this test. With the change in this PR, we'd be able to get the ptype out automatically from the recipe itself, which is great!

We would change vetiver_ptype.recipe() so it reads the value off the recipe instead of looking for it from the user in the dots.

[simonpcouch]

Documentation is super well written.

I think the fact that this PR can be so simple is a testament to how well this package is engineered. Heck yeah!

[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.