In tidymodels/butcher: Model Butcher

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(butcher)

If you come across any model objects that should be subject to butchering but does not exist in our current repository as listed here, please consider becoming a contributor to this package! For any first time contributors, this is a great place to start as we've created templates that make this process as seamless as possible.

Let's say our new model object, of class blob, was generated from an R package called "blobber". If you want to add axe methods for this class, first clone butcher onto your local computer and open up RStudio (use usethis::create_from_github("tidymodels/butcher") for an automated way to do this). After you have opened RStudio and are in the butcher RStudio project, run:

> new_model_butcher(model_class = "blob", package_name = "blobber")

You'll get the following console messages:

✔ Setting active project to 'path_to_butcher_package'
✔ Adding 'blobber' to Suggests field in DESCRIPTION
● Use `requireNamespace("blobber", quietly = TRUE)` to test if package is installed
● Then directly refer to functons like `blobber::fun()` (replacing `fun()`).

ℹ Writing skeleton files
✔ Writing 'R/blob.R'
✔ Writing 'tests/testthat/test-blob.R'
● Modify 'R/blob.R'
● Modify 'tests/testthat/test-blob.R'

new_model_butcher() leverages usethis to:

Add the new blobber modeling package under Suggests in the butcher package description file.
Generate a skeleton file under the /R directory with all possible axe methods for blob.
Generate an associated test file under /tests/testthat to test new blob axe methods.

As you can see in the R scripts for other model objects in this package, not all axe generics are always used. In fact, if you take a look at the elnet.R script, the only component of the model object fit from the package glmnet that is worth axing is the call. To help target what is worth removing from blob, we recommend first beginning with weigh() to identify which parts of the model object take up the most memory.

> weigh(fitted_blob_object)
# A tibble: 25 x 2
   object            size
   <chr>            <dbl>
 1 terms         4.01    
 2 qr.qr         0.00666 
 3 residuals     0.00286 
 4 fitted.values 0.00286 
 5 effects       0.0014  
 6 coefficients  0.00109 
 7 call          0.000728
 8 model.mpg     0.000304
 9 model.cyl     0.000304
10 model.disp    0.000304
# … with 15 more rows

In this example, the fitted model objected generated from blobber has a terms component that is taking 4.01 Mb. From here, you can examine the structure of this terms component by leveraging lobstr::sxp(fitted_blob_object$terms) or simply running utils::str(fitted_blob_object$terms). If you are looking to hunt for a specific component like the environment, fitted values, training data, controls, or the call object, take a look at locate().

Perhaps for our blob model object, we find that the call is the only piece worth axing (replacing/removing). The R/blob.R skeleton would be completed by adding a placeholder for the original call.

#' Axing a blob.
#'
#' blob model objects are created from the blobber package. They are
#' generally leveraged for classification ... insert anything relevant 
#' ... This is where all the blob specific documentation lies.
#'
#' @param x Model object.
#' @param verbose Print information each time an axe method is executed
#'  that notes how much memory is released and what functions are
#'  disabled. Default is \code{TRUE}.
#' @param ... Any additional arguments related to axing.
#'
#' @return Axed model object.
#'
#' @name axe-blob
NULL

#' Remove the call.
#'
#' @rdname axe-blob
#' @export
axe_call.blob <- function(x, verbose = TRUE, ...) {
  old <- x
  x <- exchange(x, "call", call("dummy_call"))
  if (verbose) {
    assess_object(
      old, 
      x,
      disabled = c("print", "summary")
    )
  }
  add_butcher_class(x)
}

Here we assign the current blob object x to the variable old as a means to evaluate the memory released once axe_call() is executed on the original model object. Next, we actually exchange() the current call with a dummy call of a (hopefully) smaller size. We also include assess_object() with the additional string parameter of disabled so console messages will be printed out, alerting users of any downstream functions that would be affected by axing the call. Since the original model object has different components than the new one, we add an additional butcher_blob class by calling add_butcher_class() at the end of each axe method. Once the axe methods are set, we then have a skeleton file tests/testthat/test-blob.R to aid in any unit testing.