In kenshuri/targets-tuto: Targets Tutorial

Setup

Start with a fresh data store.

library(targets)
tar_destroy()

Copy the starting _targets.R file into the working directory.

tmp <- file.copy("4-debugging/initial_targets.R", "_targets.R", overwrite = TRUE)

Open this file for editing and verify that source("4-debugging/functions.R") appears near the top.

tar_edit()

Problem

This pipeline has a bug, and your job is to find and fix it.

tar_make()

tar_visnetwork()

To begin to search for clues, review the error message and warnings stored in the metadata.

tar_meta(starts_with("run"), fields = c("error", "warnings"))

Debugging with saved workspaces

We need more information, so we invoke tar_make() with workspaces enabled. First, open _targets.R for editing.

tar_edit()

In tar_option_set(), set error to "workspace". Next time a target errors out, targets will save a workspace image file in _targets/workspaces/. (Alternatively, you could set tar_option_set(workspace = c("target1", "target2")) to always save workspaces for a specific set of targets.)

# Do not run here.
tar_option_set(
  packages = c(
    "corrr",
    "keras",
    "recipes",
    "rsample",
    "tidyverse",
    "yardstick"
  ),
  error = "workspace"
)

Now, run tar_make() and let the error happen again.

tar_make()

Confirm the name of the failed target.

library(tidyverse)
failed <- tar_meta(fields = error) %>%
  na.omit() %>%
  pull(name) %>%
  head(n = 1)
print(failed)

Confirm that a saved workspace exists for this target.

list.files("_targets/workspaces")

Load the workspace of this target.

tar_workspace(run)

The required packages, functions, RNG seed, and upstream dependencies (churn_data and churn_recipe) are now in your global environment.

ls()

Now, you can more easily reproduce the error and search for the cause.

test_model(act1 = "sigmoid", units1 = 32, churn_data, churn_recipe)

Your turn

Go find that error and fix it! You may have to explore the 4-debugging/functions.R file.

usethis::edit_file("4-debugging/functions.R")

You have completed the exercise when tar_make() runs without error. Once you are done, you may remove the workspace files either with tar_destroy(destroy = "workspaces") or unlink("_targets/workspaces/", recursive = TRUE).

Tip: workspaces option

If you want to save workspaces for a targets that do not necessarily error out, simply supply the names to the workspaces argument of tar_option_set().

Tip: interactive debugging

Base utilities debug(), debugonce(), undebug(), and browse() are extremely helpful for debugging functions. More information on interactive debugging can be found at https://rstats.wtf/debugging-r-code.html.

debugonce(test_model)
test_model(act1 = activations, units1 = units, churn_data, churn_recipe)

Once you are inside the debugger, you can access objects in the runtime environment and gradually step through the code. Press n to advance to the next line, and press c to continue on to the next breakpoint (place where the debuggger pauses). Press Q to exit the debugger.

Built-in interactive debugging in targets

As an alternative technique, you can debug interactively while tar_make() is actually running. Set these options in _targets.R.

tar_option_set(debug = "run", cue = tar_cue(mode = "never"))

Then, restart R and run tar_make(callr_function = NULL) in the R console (not in a chunk in this notebook). targets will drop you into an interactive debugger at the errored target.

# Copy and paste directly into the console.
# Do not run in this notebook. If you do, it will hang.
tar_make(callr_function = NULL)

Debugging this way is not always possible: for example, if you run tar_make_clustermq() to distribute your targets on a cluster and your target can only run on a remote compute node. However, it can often recreate the error more easily and quickly than the error = "workspace" approach.

Dynamic branching

This short course covers dynamic branching in chapter 6. Both kinds of debugging (workspaces and interactive) support dynamic branching. Interface functions like tar_option_set() can accept either individual branches or whole patterns.

kenshuri/targets-tuto documentation built on April 19, 2021, 9:58 a.m.