Defining New sl3 Learners

Introduction

This guide describes the process of implementing a learner class for a new machine learning algorithm. By writing a learner class for your favorite machine learning algorithm, you will be able to use it in all the places you could otherwise use any other sl3 learners, including Pipelines, Stacks, and Super Learner. We have done our best to streamline the process of creating new sl3 learners.

Before diving into defining a new learner, it will likely be helpful to read some background material. If you haven’t already read it, the “Modern Machine Learning in R” vignette is a good introduction to the sl3 package and it’s underlying architecture. The R6 documentation will help you understand how R6 classes are defined. In addition, the help files for sl3_Task and Lrnr_base are good resources for how those objects can be used. If you’re interested in defining learners that fit sub-learners, reading the documentation of the delayed package will be helpful.

In the following sections, we introduce and review a template for a new sl3 learner, describing the sections that can be used to define your new learner. This is followed by a discussion of the important task of documenting and testing your new learner. Finally, we conclude by explaining how you can add your learner to sl3 so that others may make use of it.

library(sl3)

Learner Template

sl3 provides a template of a learner for use in defining new learners. You can make a copy of the template to work on by invoking write_learner_template:

## Not run:
write_learner_template("path/to/write/Learner_template.R")

Let’s take a look at that template:

##' Template of a \code{sl3} Learner.
##'
##' This is a template for defining a new learner.
##' This can be copied to a new file using \code{\link{write_learner_template}}.
##' The remainder of this documentation is an example of how you might write 
##' documentation for your new learner.
##' 
##' This learner uses \code{\link[my_package]{my_ml_fun}} from \code{my_package} 
##' to fit my favorite machine learning algorithm.
##'
##' @docType class
##'
##' @importFrom R6 R6Class
##'
##' @export
##'
##' @keywords data
##'
##' @return A learner object inheriting from \code{\link{Lrnr_base}} with
##'  methods for training and prediction. For a full list of learner
##'  functionality, see the complete documentation of \code{\link{Lrnr_base}}.
##'
##' @format An \code{\link[R6]{R6Class}} object inheriting from
##'  \code{\link{Lrnr_base}}.
##'
##' @family Learners
##'
##' @section Parameters:
##'   - \code{param_1="default_1"}: This parameter does something and is not 
##'      already specified in the task. 
##'   - \code{param_2="default_2"}: This parameter does something else and is 
##'      not already specified in the task. 
##'   - \code{...}: Other parameters passed directly to 
##'      \code{\link[my_package]{my_ml_fun}}. See its documentation for details. 
##'      Also, any additional parameters that can be considered by 
##'      \code{\link{Lrnr_base}}.
##'
##' @examples 
##' include an example here, e.g. see \code{\link{Lrnr_ranger}}'s example.
##' 
Lrnr_template <- R6Class(
  classname = "Lrnr_template", 
  inherit = Lrnr_base,
  portable = TRUE, 
  class = TRUE,
  # Above, you should change Lrnr_template (in both the object name and the classname argument)
  # to a name that indicates what your learner does
  public = list(
    # you can define default parameter values here
    # if possible, your learner should define defaults for all required parameters
    initialize = function(param_1 = "default_1", param_2 = "default_2", ...) {
      # this captures all parameters to initialize and saves them as self$params
      params <- args_to_list()
      super$initialize(params = params, ...)
    },

    # you can define public functions that allow your learner to do special things here
    # for instance glm learner might return prediction standard errors
    special_function = function(arg_1) {
    }
  ),
  private = list(
    # list properties your learner supports here.
    # Use sl3_list_properties() for a list of options
    .properties = c(""),

    # .train takes task data and returns a fit object that can be used to generate predictions
    .train = function(task) {
      # generate an argument list from the parameters that were
      # captured when your learner was initialized.
      # this allows users to pass arguments directly to your ml function
      args <- self$params

      # get outcome variable type
      # preferring learner$params$outcome_type first, then task$outcome_type
      outcome_type <- self$get_outcome_type(task)
      # should pass something on to your learner indicating outcome_type
      # e.g. family or objective

      # add task data to the argument list
      # what these arguments are called depends on the learner you are wrapping
      args$x <- as.matrix(task$X_intercept)
      args$y <- outcome_type$format(task$Y)

      # only add arguments on weights and offset
      # if those were specified when the task was generated
      if (task$has_node("weights")) {
        args$weights <- task$weights
      }

      if (task$has_node("offset")) {
        args$offset <- task$offset
      }
      
      if (task$has_node("id")) {
        args$id <- task$id
      }

      # call a function that fits your algorithm
      # with the argument list you constructed
      fit_object <- call_with_args(my_ml_fun, args)

      # return the fit object, which will be stored
      # in a learner object and returned from the call
      # to learner$predict
      return(fit_object)
    },

    # .predict takes a task and returns predictions from that task
    .predict = function(task = NULL) {
      self$training_task
      self$training_outcome_type
      self$fit_object

      predictions <- predict(self$fit_object, task$X)
      return(predictions)
    },
    # list any packages required for your learner here.
    .required_packages = c("my_package")
  )
)

The template has comments indicating where details specific to the learner you’re trying to implement should be filled in. In the next section, we will discuss those details further.

Defining your Learner

Learner Name and Class

At the top of the template, we define an object Lrnr_template and set classname = "Lrnr_template". You should modify these to match the name of your new learner, which should also match the name of the corresponding R file. Note that the name should be prefixed by Lrnr_ and use snake_case.

public$initialize

This function defines the constructor for your learner, and it stores the arguments (if any) provided when a user calls make_learner(Lrnr_your_learner, ...). You can also provide default parameter values, just as the template does with param_1 = "default_1", and param_2 = "default_2". All parameters used by your newly defined learners should have defaults whenever possible. This will allow users to use your learner without having to figure out what reasonable parameter values might be. Parameter values should be documented; see the section below on documentation for details.

public$special_functions

You can of course define functions for things only your learner can do. These should be public functions like the special_function defined in the example. These should be documented; see the section below on documentation for details.

private$.properties

This field defines properties supported by your learner. This may include different outcome types that are supported, offsets and weights, amongst many other possibilities. To see a list of all properties supported/used by at least one learner, you may invoke sl3_list_properties:

sl3_list_properties()
##  [1] "binomial"      "categorical"   "continuous"    "cv"           
##  [5] "density"       "h2o"           "ids"           "importance"   
##  [9] "offset"        "preprocessing" "sampling"      "screener"     
## [13] "timeseries"    "weights"       "wrapper"

private$.required_packages

This field defines other R packages required for your learner to work properly. These will be loaded when an object of your new learner class is initialized.

User Interface for Learners

If you’ve used sl3 before, you may have noticed that while users are instructed to use learner$train, learner$predict, and learner$chain, to train, generate predictions, and generate a chained task for a given learner object, respectively, the template does not implement these methods. Instead, the template implements private methods called .train, .predict, and .chain. The specifics of these methods are explained below; however, it is helpful to first understand how the two sets of methods are related. At the risk of complicating things further, it is worth noting that there is actually a third set of methods (learner$base_train, learner$base_predict, and learner$base_chain) of which you may not be aware.

So, what happens when a user calls learner$train? That method generates a delayed object using the delayed_learner_train function, and then computes that delayed object. In turn, delayed_learner_train defines a delayed computation that calls base_train, a user-facing function that can be used to train tasks without using the facilities of the delayed package. base_train validates the user input, and in turn calls private$.train. When private$.train returns a fit_object, base_train takes that fit object, generates a learner fit object, and returns it to the user.

Each call to learner$train involves three separate training methods:

  1. The user-facing learner$train – trains a learner in a manner that can be parallelized using delayed, which calls ...
  2. ... the user-facing learner$base_train that validates user input, and which calls ...
  3. ... the internal private$.train, which does the actual work of fitting the learner and returning the fit object.

The logic in the user-facing learner$train and learner$base_train is defined in the Lrnr_base base class and is shared across all learners. As such, these methods need not be reimplemented in individual learners. By contrast, private$.train contains the behavior that is specific to each individual learner and should be reimplemented at the level of each individual learner. Since learner$base_train does not use delayed, it may be helpful to use it when debugging the training code in a new learner. The program flow used for prediction and chaining is analogous.

private$.train

This is the main training function, which takes in a task and returns a fit_object that contains all information needed to generate predictions. The fit object should not contain more data than is absolutely necessary, as including excess information will create needless inefficiencies. Many learner functions (like glm) store one or more copies of their training data – this uses unnecessary memory and will hurt learner performance for large sample sizes. Thus, these copies of the data should be removed from the fit object before it is returned. You may make use of true_obj_size to estimate the size of your fit_object. For most learners, fit_object size should not grow linearly with training sample size. If it does, and this is unexpected, please try to reduce the size of the fit_object.

Most of the time, the learner you are implementing will be fit using a function that already exists elsewhere. We’ve built some tools to facilitate passing parameter values directly to such functions. The private$.train function in the template uses a common pattern: it builds up an argument list starting with the parameter values and using data from the task, it then uses call_with_args to call my_ml_fun with that argument list. It’s not required that learners use this pattern, but it will be helpful in the common case where the learner is simply wrapping an underlying my_ml_fun.

By default, call_with_args will pass all arguments in the argument list matched by the definition of the function that it is calling. This allows the learner to silently drop irrelevant parameters from the call to my_ml_fun. Some learners either capture important arguments using dot arguments (...) or by passing important arguments through such dot arguments on to a secondary function. Both of these cases can be handled using the other_valid and keep_all options to call_with_args. The former allows you to list other valid arguments and the latter disables argument filtering altogether.

private$.predict

This is the main prediction function, and takes in a task and generates predictions for that task using the fit_object. If those predictions are 1-dimensional, they will be coerced to a vector by base_predict.

private$.chain

This is the main chaining function. It takes in a task and generates a chained task (based on the input task) using the given fit_object. If this method is not implemented, your learner will use the default chaining behavior, which is to return a new task where the covariates are defined as your learner’s predictions for the current task.

Advanced: Learners with sub-learners

Generally speaking, the above sections will be all that’s required for implementing a new learner in the sl3 framework. In some cases, it may be desirable to define learners that have “sub-learners” or other learners on which they depend. Examples of such learners are Stack, Pipeline, Lrnr_cv, and Lrnr_sl. In order to parallelize the fitting of these sub-learners using delayed, these learners implement a specialized private$.train_sublearners method that calls delayed_learner_train on their sub-learners, returning a single delayed object that, when evaluated, returns all relevant fit objects from these sub-learners. The result of that call is then passed as a second argument to their private$.train method, which now has the function prototype private$.train(task, trained_sublearners). Learners defined in such a manner usually have a much shorter computation time; the predict and chain methods are not currently parallelized in this way, although this is subject to change in the future.

If, like these learners, your learner depends on sub-learners, you have two options:

  1. Don’t worry about parallelizing sub-learners. Simply implement private$.train as discussed above, being sure to call sublearner$base_train and not sublearner$train, to avoid nesting calls to delayed, which may result in sub-optimal performance.
  2. Implement private$.train_sublearners(task) and private$.train(task, trained_sublearners), to parallelize sub-learners using delayed. We suggest reviewing the implementations of the Stack, Pipeline, Lrnr_cv and Lrnr_sl to get a better understanding of how to implement parallelized sub-learners.

In either case, you should be careful to call sublearner$base_predict and sublearner$base_chain, instead of sublearner$predict and sublearner$chain, except in the context of the private$.train_sublearners function, where you should use delayed_learner_fit_predict and delayed_learner_fit_chain.

Documenting and Testing your Learner

If you want other people to be able to use your learner, you will need to document and provide unit tests for it. The above template has example documentation, written in the roxygen format. Most importantly, you should describe what your learner does, reference any external code it uses, and document any parameters and public methods defined by it.

It’s also important to test your learner. You should write unit tests to verify that your learner can train and predict on new data, and, if applicable, generate a chained task. It might also be a good idea to use the risk function in sl3 to verify your learner’s performance on a sample dataset. That way, if you change your learner and performance drops, you know something may have gone wrong.

Submitting your Learner to sl3

Once you’ve implemented your new learner (and made sure that it has quality documentation and unit tests), please consider adding it to the sl3 project. This will make it possible for other sl3 users to use and build on your work. Make sure to add any R packages listed in .required_packages to the Suggests: field of the DESCRIPTION file of the sl3 package. Once this is done, please submit a Pull Request to the sl3 package on GitHub to request that your learned be added. If you’ve never made a “Pull Request” before, see this helpful guide: https://yangsu.github.io/pull-request-tutorial/.

From the sl3 team, thanks for your interest in extending sl3!