What is the ideal way to select dynamic targets based on upstream values?

[djbirke]

I want to run analyses of a dataset with a variety of parameters (variable definitions, grouping variables etc.) and then plot a subset of those analyses. I can put the parameters combinations into a tibble/target and then map over those to run the analyses.

But how do I filter a subset of the dynamic analysis targets based on their parameter values?

I could not find an example of this filtering step in the manual, so I came up with the following solution. (I simplified "plot" to just mean "print"). The solution is a bit unsatisfying to me, because it requires an extra target just for filtering, which possibly takes up extra storage, and because of the requirement to use with.

Is there a cleaner way to achieve my goal?

library(targets)
library(tidyverse)
library(tarchetypes)

tar_option_set(packages = "tidyverse")

list(
  tar_target(parameters, seq_len(6)),
  tar_target(analysis, tibble(parameter = parameters, output = parameters + 1), pattern = map(parameters)),
  tar_target(analysis_subset, dplyr::filter(analysis, parameter %% 2 == 1)),
  tar_target(plot, with(analysis_subset, print(output)), pattern = map(analysis_subset))
)

---

My question is closely related to ropensci/drake#1342 (comment); @wlandau writes there that switching to targets solves the storage issue. How do I achieve this?

[wlandau]

I think your workaround is great. You might also have a look at composable patterns, e.g. #212 and https://books.ropensci.org/targets/dynamic.html#pattern-construction. You might have to make deep copies of some of your branches, but if analysis_subset is small, then it might not be such a storage burden. Another option is dynamic-within-static branching, where you could statically branch over the parameter you want to filter on, dynamically branch over everything else, and just plot a single static branch.

[wlandau]

@djbirke, targets actually has a slice() pattern that I forgot to document, and it can now take multiple indexes. See #419 (comment) for a demo.

[wlandau]

See also #419 (reply in thread)

[djbirke]

Thank you a lot for your thoughts and suggestions!

If I understand correctly then your second piece of code (#406 (reply in thread)) is specific to the even/oddness of the example problem at hand, because it utilizes this trick with +1/+2 to create the static branches and then select the right subset of targets. Unfortunately, I don't think I can use such a trick in my use case.

But the first piece might work. I also read about the slice() argument. I am still trying to wrap my head around how it works. In the following, I want to convert your first code to use the slice() argument... but I can't get it right. Would you have any pointer to what I got wrong?

library(targets)

tar_script(
  {
    library(tarchetypes)
    tar_option_set(packages = "tibble")

    targets <- tar_map(
      values = list(parameter_static = seq_len(6)),
      tar_target(
        analysis,
        tibble(
          dynamic = parameter_dynamic,
          static = parameter_static,
          value = parameter_dynamic + parameter_static + 1
        ),
        pattern = map(parameter_dynamic)
      )
    )

    list(
      targets,
      tar_target(
        plot,
        print(c(analysis$dynamic, analysis$static)),
        pattern = slice(analysis, index = c(1, 3, 5))
      ),
      tar_target(parameter_dynamic, seq(10, 30, by = 10))
    )
  },
  ask = FALSE
)

tar_make()
#> ● end pipeline
#> Error : Target plot tried to branch over analysis, which is illegal. Patterns must only branch over explicitly declared targets in the pipeline. Stems and patterns are fine, but you cannot branch over branches or global objects. Also, if you branch over a target with format = "file", then that target must also be a pattern.
#> Error: callr subprocess failed: Target plot tried to branch over analysis, which is illegal. Patterns must only branch over explicitly declared targets in the pipeline. Stems and patterns are fine, but you cannot branch over branches or global objects. Also, if you branch over a target with format = "file", then that target must also be a pattern.
#> Visit https://books.ropensci.org/targets/debugging.html for debugging advice.

^{Created on 2021-04-29 by the reprex package (v1.0.0)}

[wlandau]

I think the plot target needs to go inside the tar_map() call. Also, the indexes you pass to index should be between 1 and the number of rows.

# _targets.R
library(tarchetypes)
library(targets)
tar_option_set(packages = "tibble")

targets <- tar_map(
  values = list(parameter_static = seq_len(6)),
  tar_target(
    analysis,
    tibble(
      dynamic = parameter_dynamic,
      static = parameter_static,
      value = parameter_dynamic + parameter_static + 1
    ),
    pattern = map(parameter_dynamic)
  ),
  tar_target(
    plot,
    print(c(analysis$dynamic, analysis$static)),
    pattern = slice(analysis, index = c(1, 3))
  )
)

list(
  targets,
  tar_target(parameter_dynamic, seq(10, 30, by = 10))
)

[djbirke]

Thank you, that helps a lot!