Revolutions

MLOPS for R with Azure Machine Learning

The video recording of my RStudio::conf talk, MLOPS for R with Azure Machine Learning, is now available for streaming thanks to the fine folks at RStudio.

The talk begins with a general discussion of MLOps (Machine Learning Operations) and how it differs from DevOps as applied to traditional (non-ML-based) applications. This is a theme I plan to develop further in upcoming talks, but this slide provides a summary of the main differences:

The second half of the talk describes an MLOPS workflow for building a Shiny application, based on a model trained using the caret package. I used the Azure ML service and the azuremlsdk R package to coordinate the training process and provide a cluster of machines to train multiple models simultaneously and track accuracy to choose the best model to deploy. You can find the complete code behind the demonstration shown in this vignette, included with the azuremlsdk package. (If you haven't used azuremlsdk before, start with this vignette which sets up some prerequisites.)

The slides used in the talk, along with links to other resources, are also available at the link below or at aka.ms/mlops-r.

GitHub (revodavid): Resources for Machine Learning Operations with R

Changes in the foreach package

by Hong Ooi, Senior Data Scientist at Microsoft and maintainer of the foreach package

This post is to announce some new and upcoming changes in the foreach package.

First, foreach can now be found on GitHub! The repository is at https://github.com/RevolutionAnalytics/foreach, replacing its old home on R-Forge. Right now the repo hosts both the foreach and iterators packages, but that may change later.

The latest 1.4.8 version of foreach, which is now live on CRAN, adds preliminary support for evaluating %dopar% expressions in a local environment when a sequential backend is used. This addresses a long-standing inconsistency in the behaviour of %dopar% with parallel and sequential backends, where the latter would evaluate the loop body in the global environment by default. This is a common source of bugs: code that works when prototyped with a sequential backend, mysteriously fails with a “real” parallel backend.

From version 1.4.8, the behaviour of %dopar% can be controlled with

options(foreachDoparLocal=TRUE|FALSE)

or equivalently via the system environment variable

R_FOREACH_DOPAR_LOCAL=TRUE|FALSE

with the R option taking its value from the environment variable. The current default value is FALSE, which retains the pre-existing behaviour. It is intended that over time this will be changed to TRUE.

A side-effect of this change is that %do% and %dopar% will (eventually) behave differently for a sequential backend. See this Github issue for more discussion on this topic.

In the background, the repo has also been updated to use modern tooling such as Roxygen, RMarkdown and testthat. None of these should affect how the package works, although there are some minor changes to documentation formats (in particular, the vignettes are now in HTML format rather than PDF).

Some further changes are also planned down the road, to better integrate foreach with the future package by Henrik Bengtsson. See this Github issue for further details.

Please feel free to leave comments, bug reports and pull requests at the foreach repo, or you can contact me directly at [email protected].