mlr3

Package website: release | dev

Efficient, object-oriented programming on the building blocks of machine learning. Successor of mlr.

Resources (for users and developers)

• We started writing a book. This should be the central entry point to the package.
• The mlr3gallery has some case studies and demonstrates how frequently occurring problems can be solved. It is still in early days so stay tuned for more to come.
• Videos: Currently 2 recorded talks exist from useR!2019. Both are somewhat outdated already but can be useful to get a first overview about the package ecosystem and design ideas.
  • mlr3
  • mlr3pipelines and mlr3tuning
• mlr3 reference manual
• Cheatsheets
  • Overview of cheatsheets
  • mlr3
  • mlr3tuning
  • mlr3pipelines
• Templates/Tutorials
  • mlr3-learndrake: Shows how to use mlr3 with drake for reproducible ML workflow automation.
• List of extension packages
• mlr-outreach contains public talks and slides resources.
• Our blog about mlr and mlr3. (We are not the most frequent bloggers ;) )
• Wiki: Contains mainly information for developers.

Installation

Install the last release from CRAN:

install.packages("mlr3")

Install the development version from GitHub:

remotes::install_github("mlr-org/mlr3")

Example

Constructing Learners and Tasks

library(mlr3)

# create learning task
task_iris <- TaskClassif$new(id = "iris", backend = iris, target = "Species")
task_iris

## <TaskClassif:iris> (150 x 5)
## * Target: Species
## * Properties: multiclass
## * Features (4):
##   - dbl (4): Petal.Length, Petal.Width, Sepal.Length, Sepal.Width

# load learner and set hyperparameter
learner <- lrn("classif.rpart", cp = .01)

Basic train + predict

# train/test split
train_set <- sample(task_iris$nrow, 0.8 * task_iris$nrow)
test_set <- setdiff(seq_len(task_iris$nrow), train_set)

# train the model
learner$train(task_iris, row_ids = train_set)

# predict data
prediction <- learner$predict(task_iris, row_ids = test_set)

# calculate performance
prediction$confusion

##             truth
## response     setosa versicolor virginica
##   setosa         11          0         0
##   versicolor      0         12         1
##   virginica       0          0         6

measure <- msr("classif.acc")
prediction$score(measure)

## classif.acc 
##   0.9666667

Resample

# automatic resampling
resampling <- rsmp("cv", folds = 3L)
rr <- resample(task_iris, learner, resampling)
rr$score(measure)

##             task task_id               learner    learner_id     resampling
## 1: <TaskClassif>    iris <LearnerClassifRpart> classif.rpart <ResamplingCV>
## 2: <TaskClassif>    iris <LearnerClassifRpart> classif.rpart <ResamplingCV>
## 3: <TaskClassif>    iris <LearnerClassifRpart> classif.rpart <ResamplingCV>
##    resampling_id iteration prediction classif.acc
## 1:            cv         1     <list>        0.92
## 2:            cv         2     <list>        0.92
## 3:            cv         3     <list>        0.94

rr$aggregate(measure)

## classif.acc 
##   0.9266667

Why a rewrite?

mlr was first released to CRAN in 2013. Its core design and architecture date back even further. The addition of many features has led to a feature creep which makes mlr hard to maintain and hard to extend. We also think that while mlr was nicely extensible in some parts (learners, measures, etc.), other parts were less easy to extend from the outside. Also, many helpful R libraries did not exist at the time mlr was created, and their inclusion would result in non-trivial API changes.

Design principles

• Only the basic building blocks for machine learning are implemented in this package.
• Focus on computation here. No visualization or other stuff. That can go in extra packages.
• Overcome the limitations of R’s S3 classes with the help of R6.
• Embrace R6 for a clean OO-design, object state-changes and reference semantics. This might be less “traditional R”, but seems to fit mlr nicely.
• Embrace data.table for fast and convenient data frame computations.
• Combine data.table and R6, for this we will make heavy use of list columns in data.tables.
• Defensive programming and type safety. All user input is checked with checkmate. Return types are documented, and mechanisms popular in base R which “simplify” the result unpredictably (e.g., sapply() or drop argument in [.data.frame) are avoided.
• Be light on dependencies. mlr3 requires the following packages at runtime:
  • future.apply: Resampling and benchmarking is parallelized with the future abstraction interfacing many parallel backends.
  • backports: Ensures backward compatibility with older R releases. Developed by members of the mlr team. No recursive dependencies.
  • checkmate: Fast argument checks. Developed by members of the mlr team. No extra recursive dependencies.
  • mlr3misc: Miscellaneous functions used in multiple mlr3 extension packages. Developed by the mlr team. No extra recursive dependencies.
  • paradox: Descriptions for parameters and parameter sets. Developed by the mlr team. No extra recursive dependencies.
  • R6: Reference class objects. No recursive dependencies.
  • data.table: Extension of R’s data.frame. No recursive dependencies.
  • digest: Hash digests. No recursive dependencies.
  • uuid: Create unique string identifiers. No recursive dependencies.
  • lgr: Logging facility. No extra recursive dependencies.
  • mlr3measures: Performance measures. No extra recursive dependencies.
  • mlbench: A collection of machine learning data sets. No dependencies.
• Reflections: Objects are queryable for properties and capabilities, allowing you to program on them.
• Additional functionality that comes with extra dependencies:
  • To capture output, warnings and exceptions, evaluate and callr can be used.

Extension Packages

Consult the wiki for short descriptions and links to the respective repositories.

Contributing to mlr3

This R package is licensed under the LGPL-3. If you encounter problems using this software (lack of documentation, misleading or wrong documentation, unexpected behaviour, bugs, …) or just want to suggest features, please open an issue in the issue tracker. Pull requests are welcome and will be included at the discretion of the maintainers.

Please consult the wiki for a style guide, a roxygen guide and a pull request guide.

Citing mlr3

If you use mlr3, please cite our JOSS article:

@Article{mlr3,
  title = {{mlr3}: A modern object-oriented machine learning framework in {R}},
  author = {Michel Lang and Martin Binder and Jakob Richter and Patrick Schratz and Florian Pfisterer and Stefan Coors and Quay Au and Giuseppe Casalicchio and Lars Kotthoff and Bernd Bischl},
  journal = {Journal of Open Source Software},
  year = {2019},
  month = {dec},
  doi = {10.21105/joss.01903},
  url = {https://joss.theoj.org/papers/10.21105/joss.01903},
}