In ropensci/stantargets: Targets for Stan Workflows

# With the root.dir option below,
# this vignette runs the R code in a temporary directory
# so new files are written to temporary storage
# and not the user's file space.
knitr::opts_knit$set(root.dir = fs::dir_create(tempfile()))
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
if (identical(Sys.getenv("NOT_CRAN", unset = "false"), "false")) {
  knitr::opts_chunk$set(eval = FALSE)
}
library(cmdstanr)
library(dplyr)
library(targets)
library(stantargets)
if (identical(Sys.getenv("IN_PKGDOWN"), "true")) {
  cmdstanr::install_cmdstan()
}

The stantargets package makes it easy to run a single Stan model and keep track of the results. cmdstanr fits the models, and targets manages the workflow and helps avoid unnecessary computation.

First, write a Stan model file.

lines <- "data {
  int <lower = 1> n;
  vector[n] x;
  vector[n] y;
  real true_beta;
}
parameters {
  real beta;
}
model {
  y ~ normal(x * beta, 1);
  beta ~ normal(0, 1);
}"
writeLines(lines, "x.stan")

A typical workflow proceeds as follows:

Prepare a list of input data to Stan, including vector elements x and y.
Fit the Stan model using the list of input data.
Use the fitted model object to compute posterior summaries and convergence diagnostics.
Use the fitted model object to extract posterior draws of parameters and store them in a tidy data frame.
Use the fitted model to compute Hamiltonian Monte Carlo (HMC) diagnostics.

stantargets expresses this workflow using the tar_stan_mcmc() function. To use it in a targets pipeline, invoke it from the _targets.R script of the project.

library(targets)
tar_script({
  library(stantargets)
  options(crayon.enabled = FALSE)
  tar_option_set(memory = "transient", garbage_collection = TRUE)
  generate_data <- function(n = 10) {
    true_beta <- stats::rnorm(n = 1, mean = 0, sd = 1)
    x <- seq(from = -1, to = 1, length.out = n)
    y <- stats::rnorm(n, x * true_beta, 1)
    list(n = n, x = x, y = y, true_beta = true_beta)
  }
  list(
    tar_stan_mcmc(
      example,
      "x.stan",
      generate_data(),
      stdout = R.utils::nullfile(),
      stderr = R.utils::nullfile()
    )
  )
})

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

generate_data <- function(n = 10) {
  true_beta <- stats::rnorm(n = 1, mean = 0, sd = 1)
  x <- seq(from = -1, to = 1, length.out = n)
  y <- stats::rnorm(n, x * true_beta, 1)
  list(n = n, x = x, y = y, true_beta = true_beta)
}

# The _targets.R file ends with a list of target objects
# produced by stantargets::tar_stan_mcmc(), targets::tar_target(), or similar.
list(
  tar_stan_mcmc(
    example,
    "x.stan",
    generate_data(),
    stdout = R.utils::nullfile(),
    stderr = R.utils::nullfile()
  )
)

Above, tar_stan_mcmc(example, ...) only defines the pipeline. It does not actually run Stan, it declares the targets that will eventually run Stan. Run tar_manifest() to show specific details about the targets.

tar_manifest()

Each target listed above is responsible for a piece of the workflow.

example_file_x: Reproducibly track changes to the Stan model file.
example_data: Run the code you supplied to the data argument of tar_stan_mcmc() and return a dataset compatible with Stan.
example_mcmc_x: Run the MCMC and return an object of class CmdStanMCMC.
example_draws_X: Return a friendly tibble of the posterior draws from example. Uses the $draws() method. Suppress with draws = FALSE in tar_stan_mcmc().
example_summaries_x: Return a friendly tibble of the posterior summaries from example. Uses the $summary() method. Suppress with summary = FALSE in tar_stan_mcmc().
example_diagnostics_x: Return a friendly tibble of the sampler diagnostics from example. Uses the $sampler_diagnostics() method. Suppress with diagnostics = FALSE in tar_stan_mcmc().

The suffix _x comes from the base name of the model file, in this case x.stan. If you supply multiple model files to the stan_files argument, all the models share the same dataset, and the suffixes distinguish among the various targets.

targets produces a graph to show the dependency relationships among the targets. Below, the MCMC depends on the model file and the data, and the draws, summary, and diagnostics depend on the MCMC.

tar_visnetwork(targets_only = TRUE)

Run the computation with tar_make().

tar_make()

The output lives in a special folder called _targets/ and you can retrieve it with functions tar_load() and tar_read() (from targets).

tar_read(example_summary_x)

At this point, all our results are up to date because their dependencies did not change.

tar_make()

But if we change the underlying code or data, some of the targets will no longer be valid, and they will rerun during the next tar_make(). Below, we change the Stan model file, so the MCMC reruns while the data is skipped. This behavior saves time and enhances reproducibility.

write(" ", file = "x.stan", append = TRUE)

tar_outdated()

tar_visnetwork(targets_only = TRUE)

tar_make()

At this point, we can add more targets and custom functions for additional post-processing.

library(targets)
tar_script({
  library(stantargets)
  options(crayon.enabled = FALSE)
  tar_option_set(memory = "transient", garbage_collection = TRUE)
  generate_data <- function(n = 10) {
    true_beta <- stats::rnorm(n = 1, mean = 0, sd = 1)
    x <- seq(from = -1, to = 1, length.out = n)
    y <- stats::rnorm(n, x * true_beta, 1)
    list(n = n, x = x, y = y, true_beta = true_beta)
  }
  list(
    tar_stan_mcmc(
      example,
      "x.stan",
      generate_data(),
      stdout = R.utils::nullfile(),
      stderr = R.utils::nullfile()
    ),
    tar_stan_summary(
      custom_summary,
      fit = example_mcmc_x,
      summaries = list(~posterior::quantile2(.x, probs = c(0.25, 0.75)))
    )
  )
})

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

generate_data <- function(n = 10) {
  true_beta <- stats::rnorm(n = 1, mean = 0, sd = 1)
  x <- seq(from = -1, to = 1, length.out = n)
  y <- stats::rnorm(n, x * true_beta, 1)
  list(n = n, x = x, y = y, true_beta = true_beta)
}

list(
  tar_stan_mcmc(
    example,
    "x.stan",
    generate_data(),
    stdout = R.utils::nullfile(),
    stderr = R.utils::nullfile()
  ),
  tar_stan_summary(
    custom_summary,
    fit = example_mcmc_x,
    summaries = list(~posterior::quantile2(.x, probs = c(0.25, 0.75)))
  )
)

In the graph, our new custom_summary target should be connected to the upstream example target, and only custom_summary should be out of date.

tar_visnetwork(targets_only = TRUE)

In the next tar_make(), we skip the expensive MCMC and just run the custom summary.

tar_make()

tar_read(custom_summary)

Multiple models

tar_stan_mcmc() and related functions allow you to supply multiple models to stan_files. If you do, each model will run on the same dataset. Consider a new model y.stan.

lines <- "data {
  int <lower = 1> n;
  vector[n] x;
  vector[n] y;
  real true_beta;
}
parameters {
  real beta;
}
model {
  y ~ normal(x * x * beta, 1); // Regress on x^2 instead of x.
  beta ~ normal(0, 1);
}"
writeLines(lines, "y.stan")

To include this y.stan, we add it to the stan_files argument of tar_stan_mcmc().

library(targets)
file.copy("x.stan", "y.stan")
tar_script({
  library(stantargets)
  options(crayon.enabled = FALSE)
  tar_option_set(memory = "transient", garbage_collection = TRUE)
  generate_data <- function(n = 10) {
    true_beta <- stats::rnorm(n = 1, mean = 0, sd = 1)
    x <- seq(from = -1, to = 1, length.out = n)
    y <- stats::rnorm(n, x * true_beta, 1)
    list(n = n, x = x, y = y, true_beta = true_beta)
  }
  list(
    tar_stan_mcmc(
      example,
      c("x.stan", "y.stan"), # another model
      generate_data(),
      stdout = R.utils::nullfile(),
      stderr = R.utils::nullfile()
    ),
    tar_stan_summary(
      custom_summary,
      fit = example_mcmc_x,
      summaries = list(~posterior::quantile2(.x, probs = c(0.25, 0.75)))
    )
  )
})

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

generate_data <- function(n = 10) {
  true_beta <- stats::rnorm(n = 1, mean = 0, sd = 1)
  x <- seq(from = -1, to = 1, length.out = n)
  y <- stats::rnorm(n, x * true_beta, 1)
  list(n = n, x = x, y = y, true_beta = true_beta)
}

list(
  tar_stan_mcmc(
    example,
    c("x.stan", "y.stan"), # another model
    generate_data(),
    stdout = R.utils::nullfile(),
    stderr = R.utils::nullfile()
  ),
  tar_stan_summary(
    custom_summary,
    fit = example_mcmc_x,
    summaries = list(~posterior::quantile2(.x, probs = c(0.25, 0.75)))
  )
)

In the graph below, notice how the *_x targets and *_y targets are both connected to example_data upstream.

tar_visnetwork(targets_only = TRUE)

Generated quantities

It is possible to use the CmdStanMCMC object from one run to simulate generated quantities downstream. For example, the tar_stan_gq_rep_summaries() function takes a single CmdStanMCMC object, produces multiple replications of generated quantities from multiple models, and aggregates the summaries from each. The following pipeline uses this technique to repeatedly draw from the posterior predictive distribution.

lines <- "data {
  int <lower = 1> n;
  vector[n] x;
  vector[n] y;
}
parameters {
  real beta;
}
model {
  y ~ normal(x * beta, 1);
  beta ~ normal(0, 1);
}
generated quantities {
  array[n] real y_rep = normal_rng(x * beta, 1); // posterior predictive draws
}"
writeLines(lines, "gen.stan")

library(targets)
tar_script({
  library(stantargets)
  options(crayon.enabled = FALSE)
  tar_option_set(memory = "transient", garbage_collection = TRUE)
  generate_data <- function(n = 10) {
    true_beta <- stats::rnorm(n = 1, mean = 0, sd = 1)
    x <- seq(from = -1, to = 1, length.out = n)
    y <- stats::rnorm(n, x * true_beta, 1)
    list(n = n, x = x, y = y, true_beta = true_beta)
  }
  list(
    tar_stan_mcmc(
      example,
      "x.stan",
      generate_data(),
      stdout = R.utils::nullfile(),
      stderr = R.utils::nullfile()
    ),
    tar_stan_gq_rep_summary(
      postpred,
      stan_files = "gen.stan",
      fitted_params = example_mcmc_x, # one CmdStanFit object
      data = generate_data(), # multiple simulated datasets
      batches = 2, # 2 dynamic branches
      reps = 5, # 5 replications per branch
      quiet = TRUE,
      stdout = R.utils::nullfile(),
      stderr = R.utils::nullfile()
    )
  )
})

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

generate_data <- function(n = 10) {
  true_beta <- stats::rnorm(n = 1, mean = 0, sd = 1)
  x <- seq(from = -1, to = 1, length.out = n)
  y <- stats::rnorm(n, x * true_beta, 1)
  list(n = n, x = x, y = y, true_beta = true_beta)
}

list(
  tar_stan_mcmc(
    example,
    "x.stan",
    generate_data(),
    stdout = R.utils::nullfile(),
    stderr = R.utils::nullfile()
  ),
  tar_stan_gq_rep_summary(
    postpred,
    stan_files = "gen.stan",
    fitted_params = example_mcmc_x, # one CmdStanFit object
    data = generate_data(), # Function runs once per rep.
    batches = 2, # 2 dynamic branches
    reps = 5, # 5 replications per branch
    quiet = TRUE,
    stdout = R.utils::nullfile(),
    stderr = R.utils::nullfile()
  )
)