Skip to contents

Running bayesim with targets

Source: vignettes/targets.Rmd
targets.Rmd

Introduction

targets is a Make-like pipeline engine for R: it caches the return value of each step and only re-runs a step when its inputs change. That maps onto bayesim naturally because every simulation has a stable config fingerprint — a SHA256 hash of the data/fit grids, the generator/fitter/metrics specs, n_replicates, and the seed (see vignette("reproducibility")). The exported config_fingerprint() accessor is useful as an audit ID and as a compact dependency for downstream targets.

targets is not a hard dependency of bayesim. The pipeline code below is eval = FALSE; install targets separately if you want to run it.

A minimal targets pipeline

A typical workflow defines one dynamically branched target per simulation condition (so conditions rebuild independently) or one target for the whole config. targets already hashes commands and dependencies under its default tar_cue(mode = "thorough"); there is no fingerprint argument to tar_cue(). The pipeline below relies on that native invalidation and records bayesim’s fingerprint as a separate target.

# _targets.R
library(targets)
library(bayesim)

# Build the data and fit grids once, up front.
data_grid <- data.frame(
  n = c(100, 500, 1000),
  effect = c(0.2, 0.5, 1.0)
)
fit_grid <- data.frame(model = c("baseline", "full"))

# One row per condition; we will map over it to define one target per condition.
conditions <- split(expand.grid(
  data_idx = seq_len(nrow(data_grid)),
  fit_idx = seq_len(nrow(fit_grid)),
  stringsAsFactors = FALSE
), seq_len(nrow(data_grid) * nrow(fit_grid)))

list(
  tar_target(condition, conditions, iteration = "list"),

  # One dynamic branch per (data_idx, fit_idx) condition.
  tar_target(
    sim_condition,
    {
      config <- simulation_config(
        data_grid = data_grid[condition$data_idx, , drop = FALSE],
        fit_grid = fit_grid[condition$fit_idx, , drop = FALSE],
        data_generator = my_data_generator,
        fitter = LinearRegressionFitter(n_draws = 500L),
        metrics = list(posterior_summary_metric(), coverage_metric()),
        n_replicates = 200L,
        seed = 42L
      )
      run_simulation(config, progress = FALSE)
    },
    pattern = map(condition)
  ),

  # A single-target alternative for the whole study.
  tar_target(
    config_all,
    simulation_config(
      data_grid = data_grid,
      fit_grid = fit_grid,
      data_generator = my_data_generator,
      fitter = LinearRegressionFitter(n_draws = 500L),
      metrics = list(posterior_summary_metric(), coverage_metric()),
      n_replicates = 200L,
      seed = 42L
    )
  ),
  tar_target(config_id, config_fingerprint(config_all)),
  tar_target(
    sim_all,
    {
      config_id # explicit audit dependency, also visible in the target graph
      run_simulation(config_all, progress = FALSE)
    }
  )
)

Notes on invalidation:

  • The default thorough cue rebuilds a target when its command or upstream dependencies change. No custom cue is needed.
  • config_id excludes runtime policy (result_path, retention, checkpoint settings, max_errors, and stop_on) by design, making it a stable study ID. sim_all still depends on the full config_all, so changing runtime policy can correctly affect execution and checkpoint behavior.

Combining results

Once the per-condition targets have built, assemble them into a single performance table with tar_read() / tar_load(). bayesim’s performance_measures() expects a summary data frame, so we row-bind the per-condition summaries first:

# In an interactive session, after tar_make():
library(targets)
# Load all per-condition results.
conditions_results <- tar_read(sim_condition)          # list, one element per branch
combined <- do.call(rbind, lapply(conditions_results, \(r) r$summary))

pm <- performance_measures(combined, estimand = "effect")
pm

For the whole-config target, tar_read(sim_all)$summary gives the full per-task summary directly, and summarize_simulation(tar_read(sim_all)) produces the aggregated condition-level table with MCSEs.

Next steps