How does CmdStanR work?

Introduction

This vignette is intended to be read after the Getting started with CmdStanR vignette. Please read that first for important background. In this document we provide additional details about compiling models, passing in data, and how CmdStan output is saved and read back into R.

We will only use the $sample() method in examples, but all model fitting methods work in a similar way under the hood.

library(cmdstanr)
check_cmdstan_toolchain(fix = TRUE, quiet = TRUE)

Compilation

Immediate compilation

The cmdstan_model() function creates a new CmdStanModel object. The CmdStanModel object stores the path to a Stan program as well as the path to a compiled executable.

stan_file <- file.path(cmdstan_path(), "examples", "bernoulli", "bernoulli.stan")
mod <- cmdstan_model(stan_file)
mod$print()
mod$stan_file()
mod$exe_file()

Subsequently, if you create a CmdStanModel object from the same Stan file then compilation will be skipped (assuming the file hasn’t changed).

mod <- cmdstan_model(stan_file)

Internally, cmdstan_model() first creates the CmdStanModel object from just the Stan file and then calls its $compile() method. Optional arguments to the $compile() method can be passed via ....

mod <- cmdstan_model(
  stan_file,
  force_recompile = TRUE,
  include_paths = "paths/to/directories/with/included/files",
  cpp_options = list(stan_threads = TRUE, STANC2 = TRUE)
)

Delayed compilation

It is also possible to delay compilation when creating the CmdStanModel object by specifying compile=FALSE and then later calling the $compile() method directly.

unlink(mod$exe_file())
mod <- cmdstan_model(stan_file, compile = FALSE)
mod$exe_file() # not yet created
mod$compile()
mod$exe_file()

Pedantic check

If you are using CmdStan version 2.24 or later and CmdStanR version 0.2.1 or later, you can run a pedantic check for your model. CmdStanR will always check that your Stan program does not contain any invalid syntax but with pedantic mode enabled the check will also warn you about other potential issues in your model, for example:

  • Distribution usages issues: distribution arguments do not match the distribution specification, or some specific distribution is used in an inadvisable way.
  • Unused parameter: a parameter is defined but does not contribute to target.
  • Large or small constant in a distribution: very large or very small constants are used as distribution arguments.
  • Control flow depends on a parameter: branching control flow (like if/else) depends on a parameter value.
  • Parameter has multiple twiddles: a parameter is on the left-hand side of multiple twiddles (i.e., multiple ~ symbols).
  • Parameter has zero or multiple priors: a parameter has zero or more than one prior distribution.
  • Variable is used before assignment: a variable is used before being assigned a value.
  • Strict or nonsensical parameter bounds: a parameter is given questionable bounds.

For the latest information on the checks performed in pedantic mode see the Pedantic mode chapter in the Stan Reference Manual.

Pedantic mode is available when compiling the model or when using the separate $check_syntax() method of a CmdStanModel object. Internally this corresponds to setting the stanc (Stan transpiler) option warn-pedantic. Here we demonstrate pedantic mode with a Stan program that is syntactically correct but is missing a lower bound and a prior for a parameter.

stan_file_pedantic <- write_stan_file("
data {
  int N;
  array[N] int y;
}
parameters {
  // should have <lower=0> but omitting to demonstrate pedantic mode
  real lambda;
}
model {
  y ~ poisson(lambda);
}
")

To turn on pedantic mode at compile time you can set pedantic=TRUE in the call to cmdstan_model() (or when calling the $compile() method directly if using the delayed compilation approach described above).

mod_pedantic <- cmdstan_model(stan_file_pedantic, pedantic = TRUE)

To turn on pedantic mode separately from compilation use the pedantic argument to the $check_syntax() method.

mod_pedantic$check_syntax(pedantic = TRUE)

Using pedantic=TRUE via the $check_syntax() method also has the advantage that it can be used even if the model hasn’t been compiled yet. This can be helpful because the pedantic and syntax checks themselves are much faster than compilation.

file.remove(mod_pedantic$exe_file()) # delete compiled executable
rm(mod_pedantic)

mod_pedantic <- cmdstan_model(stan_file_pedantic, compile = FALSE)
mod_pedantic$check_syntax(pedantic = TRUE)

Stan model variables

If using CmdStan 2.27 or newer, you can obtain the names, types and dimensions of the data, parameters, transformed parameters and generated quantities variables of a Stan model using the $variables() method of the CmdStanModel object.

stan_file_variables <- write_stan_file("
data {
  int<lower=1> J;
  vector<lower=0>[J] sigma;
  vector[J] y;
}
parameters {
  real mu;
  real<lower=0> tau;
  vector[J] theta_raw;
}
transformed parameters {
  vector[J] theta = mu + tau * theta_raw;
}
model {
  target += normal_lpdf(tau | 0, 10);
  target += normal_lpdf(mu | 0, 10);
  target += normal_lpdf(theta_raw | 0, 1);
  target += normal_lpdf(y | theta, sigma);
}
")
mod_v <- cmdstan_model(stan_file_variables)
variables <- mod_v$variables()

The $variables() method returns a list with data, parameters, transformed_parameters and generated_quantities elements, each corresponding to variables in their respective block of the program. Transformed data variables are not listed as they are not used in the model’s input or output.

names(variables)
names(variables$data)
names(variables$parameters)
names(variables$transformed_parameters)
names(variables$generated_quantities)

Each variable is represented as a list containing the type information (currently limited to real or int) and the number of dimensions.

variables$data$J
variables$data$sigma
variables$parameters$tau
variables$transformed_parameters$theta

Executable location

By default, the executable is created in the same directory as the file containing the Stan program. You can also specify a different location with the dir argument.

mod <- cmdstan_model(stan_file, dir = "path/to/directory/for/executable")

Processing data

There are three data formats that CmdStanR allows when fitting a model:

  • named list of R objects
  • JSON file
  • R dump file

Named list of R objects

Like the RStan interface, CmdStanR accepts a named list of R objects where the names correspond to variables declared in the data block of the Stan program. In the Bernoulli model the data is N, the number of data points, and y an integer array of observations.

mod$print()
# data block has 'N' and 'y'
data_list <- list(N = 10, y = c(0,1,0,0,0,0,0,0,0,1))
fit <- mod$sample(data = data_list)

Because CmdStan doesn’t accept lists of R objects, CmdStanR will first write the data to a temporary JSON file using write_stan_json(). This happens internally, but it is also possible to call write_stan_json() directly.

data_list <- list(N = 10, y = c(0,1,0,0,0,0,0,0,0,1))
json_file <- tempfile(fileext = ".json")
write_stan_json(data_list, json_file)
cat(readLines(json_file), sep = "\n")

JSON file

If you already have your data in a JSON file you can just pass that file directly to CmdStanR instead of using a list of R objects. For example, we could pass in the JSON file we created above using write_stan_json():

fit <- mod$sample(data = json_file)

R dump file

Finally, it is also possible to use the R dump file format. This is not recommended because CmdStan can process JSON faster than R dump, but CmdStanR allows it because CmdStan will accept files created by rstan::stan_rdump():

rdump_file <- tempfile(fileext = ".data.R")
rstan::stan_rdump(names(data_list), file = rdump_file, envir = list2env(data_list))
cat(readLines(rdump_file), sep = "\n")
fit <- mod$sample(data = rdump_file)

Writing CmdStan output to CSV

Default temporary files

data_list <- list(N = 10, y = c(0,1,0,0,0,0,0,0,0,1))
fit <- mod$sample(data = data_list)

When fitting a model, the default behavior is to write the output from CmdStan to CSV files in a temporary directory.

fit$output_files()

These files will be lost if you end your R session or if you remove the fit object and force (or wait for) garbage collection.

files <- fit$output_files()
file.exists(files)

rm(fit)
gc()

file.exists(files)

Non-temporary files

To save these files to a non-temporary location there are two options. You can either specify the output_dir argument to mod$sample() or use fit$save_output_files() after fitting the model.

# see ?save_output_files for info on optional arguments
fit$save_output_files(dir = "path/to/directory")
fit <- mod$sample(
  data = data_list,
  output_dir = "path/to/directory"
)

Reading CmdStan output into R

Lazy CSV reading

With the exception of some diagnostic information, the CSV files are not read into R until their contents are requested by calling a method that requires them (e.g., fit$draws(), fit$summary(), etc.). If we examine the structure of the fit object, notice how the Private slot draws_ is NULL, indicating that the CSV files haven’t yet been read into R.

str(fit)

After we call a method that requires the draws then if we reexamine the structure of the object we will see that the draws_ slot in Private is no longer empty.

draws <- fit$draws() # force CSVs to be read into R
str(fit)

For models with many parameters, transformed parameters, or generated quantities, if only some are requested (e.g., by specifying the variables argument to fit$draws()) then CmdStanR will only read in the requested variables (unless they have already been read in).

read_cmdstan_csv()

Internally, the read_cmdstan_csv() function is used to read the CmdStan CSV files into R. This function is exposed to users, so you can also call it directly.

# see ?read_cmdstan_csv for info on optional arguments controlling
# what information is read in
csv_contents <- read_cmdstan_csv(fit$output_files())
str(csv_contents)

as_cmdstan_fit()

If you need to manually create fitted model objects from CmdStan CSV files use as_cmdstan_fit().

fit2 <- as_cmdstan_fit(fit$output_files())

This is pointless in our case since we have the original fit object, but this function can be used to create fitted model objects (CmdStanMCMC, CmdStanMLE, etc.) from any CmdStan CSV files.

Saving and accessing advanced algorithm info (latent dynamics)

If save_latent_dynamics is set to TRUE when running the $sample() method then additional CSV files are created (one per chain) that provide access to quantities used under the hood by Stan’s implementation of dynamic Hamiltonian Monte Carlo.

CmdStanR does not yet provide a special method for processing these files but they can be read into R using R’s standard CSV reading functions.

fit <- mod$sample(data = data_list, save_latent_dynamics = TRUE)
fit$latent_dynamics_files()

# read one of the files in
x <- utils::read.csv(fit$latent_dynamics_files()[1], comment.char = "#")
head(x)

The column lp__ is also provided via fit$draws(), and the columns accept_stat__, stepsize__, treedepth__, n_leapfrog__, divergent__, and energy__ are also provided by fit$sampler_diagnostics(), but there are several columns unique to the latent dynamics file.

head(x[, c("theta", "p_theta", "g_theta")])

Our model has a single parameter theta and the three columns above correspond to theta in the unconstrained space (theta on the constrained space is accessed via fit$draws()), the auxiliary momentum p_theta, and the gradient g_theta. In general, each of these three columns will exist for every parameter in the model.

Developing using CmdStanR

CmdStanR can of course be used for developing other packages that require compiling and running Stan models as well as using new or custom Stan features available through CmdStan.

Pre-compiled Stan models in R packages

You may compile a Stan model at runtime (e.g. just before sampling), or you may compile all the models inside the package file system in advance at installation time. The latter avoids compilations at runtime, which matters in centrally managed R installations where users should not compile their own software.

To pre-compile all the models in a package, you may create top-level scripts configure and configure.win which run cmdstan_model() with compile = TRUE and save the compiled executables somewhere inside the inst/ folder of the package source. The instantiate package helps developers configure packages this way, and it documents other topics such as submitting to CRAN and administering CmdStan. Kevin Ushey’s configure package helps create and manage package configuration files in general.

Troubleshooting and debugging

When developing or testing new features it might be useful to have more information on how CmdStan is called internally and to see more information printed when compiling or running models. This can be enabled for an entire R session by setting the option "cmdstanr_verbose" to TRUE.

options("cmdstanr_verbose"=TRUE)

mod <- cmdstan_model(stan_file, force_recompile = TRUE)
fit <- mod$sample(
  data = data_list,
  chains = 1,
  iter_warmup = 100,
  iter_sampling = 100
)