Should we change the underlying mechanism that creates the specifications to call internal functions?

[spsanderson]

Should a separate function be built for each model sort of like so?

> library(tidyAML)
> library(dplyr)
> 
> lm_regression_linear_reg <- function(.pe, .pm) {
+   parsnip::linear_reg() |>
+     parsnip::set_mode(.pm) |>
+     parsnip::set_engine(.pe)
+ }
> 
> mod_tbl <- make_regression_base_tbl() |>
+   slice(1) |>
+   mutate(
+     fns_call = paste(.parsnip_engine, .parsnip_mode, .parsnip_fns, sep = "_")
+   ) |>
+   mutate(
+     .mod_spec = list(match.fun(fns_call)(.parsnip_engine, .parsnip_mode))
+   ) |>
+   select(-fns_call)
> 
> mod_tbl
# A tibble: 1 × 4
  .parsnip_engine .parsnip_mode .parsnip_fns .mod_spec
  <chr>           <chr>         <chr>        <list>   
1 lm              regression    linear_reg   <spec[+]>
> mod_tbl[[".mod_spec"]][[1]]
Linear Regression Model Specification (regression)

Computational engine: lm 

Currently the function make_regression_base_tbl() is used to get the following:

> tidyAML::make_regression_base_tbl()
# A tibble: 39 × 3
   .parsnip_engine .parsnip_mode .parsnip_fns
   <chr>           <chr>         <chr>       
 1 lm              regression    linear_reg  
 2 brulee          regression    linear_reg  
 3 gee             regression    linear_reg  
 4 glm             regression    linear_reg  
 5 glmer           regression    linear_reg  
 6 glmnet          regression    linear_reg  
 7 gls             regression    linear_reg  
 8 lme             regression    linear_reg  
 9 lmer            regression    linear_reg  
10 stan            regression    linear_reg  
# ℹ 29 more rows
# ℹ Use `print(n = ...)` to see more rows

This creates the base specification before the actual model specs are built. Right now in practice purrr::imap is used to create the specifications like so:

function(.model_tbl){

  # Variables ----
  model_tbl <- .model_tbl

  # Checks ----
  if (!inherits(model_tbl, "tidyaml_base_tbl")){
    rlang::abort(
      message = "The model tibble must come from the make base tbl function.",
      use_cli_format = TRUE
    )
  }

  # Manipulation
  model_factor_tbl <- model_tbl |>
    dplyr::mutate(.model_id = dplyr::row_number() |>
                    forcats::as_factor()) |>
    dplyr::select(.model_id, dplyr::everything())

  # Make a group split object list
  models_list <- model_factor_tbl |>
    dplyr::group_split(.model_id)

  # Make the Workflow Object using purrr imap
  model_spec <- models_list |>
    purrr::imap(
      .f = function(obj, id){

        # Pull the model column and then pluck the model
        pe <- obj |> dplyr::pull(2) |> purrr::pluck(1)
        pm <- obj |> dplyr::pull(3) |> purrr::pluck(1)
        pf <- obj |> dplyr::pull(4) |> purrr::pluck(1)

        ret <- match.fun(pf)(mode = pm, engine = pe)

        # Return the result
        attributes(ret)$.tidyaml_mod_class <- paste0(base::tolower(pe), "_", base::tolower(pf))
        return(ret)
      }
    )

  # Return
  # Make sure to return as a tibble
  model_spec_ret <- model_factor_tbl |>
    dplyr::mutate(model_spec = model_spec)  |>
    dplyr::mutate(.model_id = as.integer(.model_id))

  return(model_spec_ret)
}

Should we explore different ways of doing this like so:

mod_mode <- "regression"

mod_tbl_b <- tibble::tribble(
    ~.parsnip_engine, ~.parsnip_mode, ~.parsnip_fns,
    "lm", "regression", parsnip::linear_reg(mode = mod_mode, engine = "lm"),
    "brulee", "regression", parsnip::linear_reg(mode = mod_mode, engine = "brulee"),
    "gee", "regression", parsnip::linear_reg(mode = mod_mode, engine = "gee")
  )

> mod_tbl_b
# A tibble: 3 × 3
  .parsnip_engine .parsnip_mode .parsnip_fns
  <chr>           <chr>         <list>      
1 lm              regression    <spec[+]>   
2 brulee          regression    <spec[+]>   
3 gee             regression    <spec[+]> 

OR

mod_tbl_c <- tibble::tribble(
    ~.parsnip_engine, ~.parsnip_mode, ~.parsnip_fns,
    "lm", "regression", parsnip::linear_reg(),
    "brulee", "regression", parsnip::linear_reg(),
    "gee", "regression", parsnip::linear_reg()
  ) |>
  dplyr::mutate(
    .parsnip_fns = purrr::map2(
      .x = .parsnip_fns,
      .y = .parsnip_engine,
      \(x, y) x |> parsnip::set_engine(y)
    )
  )

> mod_tbl_c$.parsnip_fns
[[1]]
Linear Regression Model Specification (regression)

Computational engine: lm 


[[2]]
Linear Regression Model Specification (regression)

Computational engine: brulee 


[[3]]
! parsnip could not locate an implementation for `linear_reg` model specifications using the `gee` engine.
ℹ The parsnip extension package multilevelmod implements support for this specification.
ℹ Please install (if needed) and load to continue.

Linear Regression Model Specification (regression)

Computational engine: gee