GitHub - mlr-org/mlr3-targets (original) (raw)

mlr3-targets

The goal of mlr3-targets is to show how to use the mlr3 machine-learning framework in combination with the workflow package targets.

This example project showcases a benchmark of different learners (SVM, KKNN, RF), including hyperparameter tuning, across the iris and spam datasets.

The project shows examples of

• dynamic branching in targets
• creation of custom functions
• using multiple "plans" to organize the project
• the use of {renv} for package version control

Usage

To clone this repo, execute the following

usethis::use_course("mlr-org/mlr3-targets")

To install a fixed snapshot of the required R packages call

To install the latest versions of the required R packages call

After a successful installation of all dependencies call `

to run the complete project. Alternative, use tar_make_clustermq() to run in parallel.

• This will build all R objects of the project (or "targets" in target's DSL) in the correct order.
• You can visualize the project dependency structure via tar_visnetwork().
• To load specific R objects into the global environment, call tar_load(<object name>).

See the targets manual for more information on {targets}.

Other targets learning resources

• Talks
• targets-examples
• targets

Acknowledgements

• Will Landau for developing and maintaining {targets} in an awesome way
• The mlr3 team for developing the {mlr3} package framework

Custom structure justification

This project uses a custom, personal structure for targets-based projects. The following bullet points outline the thoughts behind this structure.

• Using R.utils::sourceDirectory() instead of a for-loop to source multiple scripts/directories makes _targets.R a bit cleaner with a minimal increase WRT to dependencies
• Putting targets into plans/ (instead of all into _targets.R) and splitting them up across multiple R scripts allows for a meta-level organization of targets. Including the scripts individually in _targets.R allows to quickly comment out certain ones (which might relate to a standalone project part). This seems easier than searching for specific target which would avoid other project parts to be run
• Each "plan" is visible in the global environment including the individual target count
• The decision to place the list of required packages in packages.R instead of _targets.R is for the simple reason that the name of packages.R is very descriptive. When a new package is required, I just think "packages" in my head and grep for packages.R to add a new package.