Reproducible machine learning workflows
Patrick Schratz
September 27th, 2019
usethis::use_course("mlr-org/mlr3-learndrake")
Find someone to sit next to and share laptops
(Slide style and content borrowed from Garrick Aden-Buie )
What is drake?
What is drake?
Will this work when I come back to it later?
What happens if I re-run everything?
Am I certain that the results are still valid?
(borrowed from ropensci/drake)
drake Essentials
_drake.R
The main settings file
sources all R functions
sources all "drake-plans" and binds them together
sets the execution parameters in
drake_config()called by
r_make()at the start of every run
drake "targets"
targets are simple R objects which know their dependencies
targets represent all intermediate or final results of your project
targets live in the drake-cache (
.drakedirectory) and can from there be loaded into the global environment usingloadd()
Advantages
No need to use
saveRDS()/save()anymore to store intermediate outputsQuick-load targets under cursor in RStudio via addin
visualize dependencies via
r_vis_drake_graph()
r_vis_drake_graph()
Shows how all targets are connected
How does drake help you in your daily work?
Knows the execution order of your complete analysis
r_vis_drake_graph()Enables running your complete project in one call
r_make()Tracks metadata of each target (runtime, etc.)
If you make a substantial change downstream, all subsequent targets will be updated (plots, tables, models, etc.)
But wait, there's more!
There is a lot more that drake can do, including:
- Predict total runtime of your plan
r_predict_runtime()Support for large plans by using wildcards and functions
map(),cross()andcombine()Parallel computation of targets via packages
future
clustermq (HPC)
Support for many HPC schedulers (SLURM, SGE, PBS, LSF) via package clustermq
Learn more with these resources
Drake
Icons by Github octicons and the following Noun Project icon creators: Kirby Wu, TW, Iga, Lil Squid, Wichai Wi, Nick Kinling.
Built with rmarkdown, xaringan, xaringanthemer, remark.js, tachyons.css, animate.css
Personal FAQ (1/2)
- What is the difference between
drake::make()anddrake::r_make()?
r_make()runs in a fresh R session via package callr and sources the config before every run. No need to source the config yourself and then runmake(config). See also drake-manual/safer-interactivity.
- How do I parallelize code within targets?
By passing
prework = quote(future::plan(future.callr::callr, workers = <n>))or a similar parallel plan todrake_config()in_drake.R. Do not usefuture::plan()within scripts/functions.
- I have no idea what the status of my running computation is. Is there a way to see it?
Yes, either use
drake::progress(),drake::finished()ordrake::failed(). Optionally make the output more verbose usingdrake::progress() %>% print(n = 200).
Personal FAQ (2/2)
- Everything in the drake manual is forcing me to use functions but I am used to scripts mainly. What is better with functions?
Project structure and organization is eminently important. A function-based workflow enhances reproducibility and leads to a more clean project structure. See this discussion for more information.
- Is there a convenient way to load targets quickly?
Yes! With the installation of drake there is an addin called "load targets under cursor" that lets you load the R object under your cursor when its bound to a keyboard shortcut. Mine is bound to "CTRL + L".
- What about other addins and keyboard shortcuts?
Take a look into the "Addins" menu in RStudio for a list of avaiable addins. I have the followings addin/shorcut bindings:
r_vis_drake_graph()-> "CTRL + Shift + V";r_outdated()-> "CTRL + Shift + O"; "load target under cursor" -> "CTRL +L"