An introduction to targets for R

R-Ladies Santa Barbara

October 11, 2023

Tracey Mangin

1

we have 15 slides!

Agenda

Resources

• This presentation (hi!) ~20 minutes
• reasonable econ style, interrupt with questions (but use your best judgement)
• Follow along demonstration ~20 minutes
• Example in groups (or individually, or all together!)

• There are many resources available online!
• I pulled information from several sources - thanks! (listed at the end)
• Check those out!

2

In a perfect world…

3

import inputs

clean and process data

perform analysis

summarize/visualize outputs

we’re done hooray!

In a perfect world…

Reality…

4

import inputs

clean and process data

perform analysis

summarize/visualize outputs

import inputs

clean and process data

perform analysis

summarize/visualize outputs

update inputs

change cleaning

change this a thousand times and add a million more steps

are these right?

next time, i promise we’ll be perfect…

we’re done hooray!

Workflow challenges

1. Long run times
1. Rerunning entire pipelines to ensure that items are up to date can take a lot of time��
2. Reproducibility
• spent wayyy too much time making this

​

5

Enter targets package for R

• Pipeline tool specifically for R
• Pipeline tools coordinate the pieces of analysis projects (e.g., Make)
• Keeps track of entire workflow
• Automatically detects when files or functions change
• Saves time by only running steps, or targets, that are no longer up to date
• Ensures that the pipeline is run in the correct order
• Ensures reproducibility: When targets are up to date, this is evidence that the outputs match the code and inputs
• More trustworthy and reproducible results
• Note: targets replaces the R tool drake

6

target explained

• Each step in the pipeline = a target
• Looks and feels like a variable (name)
• Stores the returned R object in your project folder (not in the environment)
• Function oriented
• Target usually creates, analyzes, or summarizes a dataset/analysis
• Good targets:
• Large enough to save time when not run
• Small enough that some are skipped
• Don’t modify global environment
• Return a single “value” or R object

7

function

clean_data <- function(file) {

​

data <- read_csv(file) %>%

filter(!is.na(date_time)) %>%

mutate(day = weekdays(date_time)

​

data

​

}

​

target

tar_target(name = data, command = clean_data(file))

​

other examples

tar_target(name = max_val, command = 16))

​

tar_target(name = save_data, command = simple_write(data.csv), format “file”)

R code to run target

target name

targets setup

1. Create a project or repo
2. Create a folder called R
1. Save script(s) containing functions for analysis in R folder (these scripts are sourced)
3. Add input files if storing in project
4. Install targets (once): install.packages("targets")
5. Run use_targets()
• Creates required _targets.R file that runs pipeline
6. Set options (e.g., libraries)
7. Fill in list() with targets

8

Inspect pipeline once targets filled in

9

functions

tar_manifest()helps check for obvious errors and produces a data frame of info about the targets in the pipeline

​

tar_visnetwork()visualizes pipeline workflow

​

tar_glimpse()visualizes pipeline workflow faster than visnetwork, but doesn’t account for progress info

• to see functions, set targets_only = FALSE

tar_visnetwork()

tar_glimpse()

Run the pipeline

• tar_make() runs the pipeline
• Runs targets in the correct order
• Only runs targets that are out of date (time saver!)

​

• Creates folder called _targets in project
• Outputs saved in _targets/objects
• tar_read() prints the output
• tar_load() loads the output in the environment

​

10

Making changes and rerunning

• Reruns: targets identifies which parts of pipeline are outdated and only reruns those… a real time saver!
• tar_outdated() returns names of outdated targets

​

• tar_visnetwork() visualizes pipeline and shows out of date targets

11

tar_visnetwork()

Debugging

• Different because not interactive
• Layers that make targets good for reproducibility and scaling make it harder to debug
• If you have an error, run tar_make() will return error message
• Can run working parts if you set tar_option_set(error = “null”), which will return NULL for errored targets
• Note: outputs not up to date or correct, but this allows you to look at the outputs!

12

Debugging steps

• _targets/meta/meta metadata stores most recent error
• tar_meta() can retrieve error messages
• Look at functions
• Most errors are in user defined functions
• Pause pipeline with browser()
• Personal approach: step through functions as you normally would
• tar_destroy() removes that _targets folder - not best practice though!

13

More advanced info and resources

• _targets.R setup has code for parallel processing
• If running on a cluster, use_targets() would have detected this and set up _targets.R for parallel processing

CHECK THESE OUT!

1. The {targets} R package user manual
2. Get started with {targets} in 4 minutes
3. https://docs.ropensci.org/targets/
4. Will Landau - Reproducible Computation at Scale in R with Targets [Remote]
5. Reproducible Computation at Scale in R with {targets}
6. FULL TUTORIAL: Build a Full Production Forecasting Workflow in R with Targets & Modeltime
1. Includes parallelization (see table of contents)
7. R {targets}: How to Make Reproducible Pipelines for Data Science and Machine Learning - Machine Learning, R programming

14

Demo

• Data: listen history from https://www.last.fm/ (~2.5 weeks)
• Workflow
• Load input file
• Clean input
• Summarize data
• Create visualizations

15