Skip to contents

Changelog

Source: NEWS.md

CmdStanR 1.0.0

This is a major release with enough noteworthy items that we’ve broken the release notes into sections. The Building models section covers the compilation redesign. In particular, cmdstan_model() now compiles the model or reuses an up-to-date executable, before it returns, and the separate $compile() method is gone (code that calls $compile() or passes compile = FALSE will need to change). Data and initial values covers improvements to how data and initial values are handled, Other new features and Bug fixes cover what their titles suggest, and Removed and deprecated lists everything else that could break existing code beyond the compilation redesign.

Where a change can break existing code, the entry says what to use instead, so you (or your coding agent) can update it.

A massive thank you to everyone who contributed code and reviews to this release and to everyone who filed the bug reports and feature requests that went into it.

Building models

  • cmdstan_model() now compiles the model, or reuses an up-to-date executable, before it returns. The compile argument and the $compile() method are gone. To compile a Stan file without creating a model object, use the new compile_stan_file(), which returns the path to the executable. The new check_syntax_stan_file(), format_stan_file() and variables_stan_file() also work without a model object. (#1256)
  • cmdstan_model() now recompiles when the Stan program or its #includes, the user header, cpp_options, stanc_options, make/local or the CmdStan installation has changed, and in an interactive session (or with quiet = FALSE) it says which. For changes it can’t detect, such as a new C++ compiler, use force_recompile = TRUE. The vignette “How does CmdStanR work?” lists them. (#1255, #1237, #1019, #1308)
  • Compiling now writes a hidden file next to the executable that records how it was built, for example .bernoulli.cmdstanr.json next to bernoulli. If your .gitignore ignores the executable, add .*.cmdstanr.json too. If the file is missing, cmdstan_model() recompiles. (#1238)
  • The new stan_build_info() and $build_info() show how an executable was built: the options and files used, the CmdStan installation, and what the executable itself reports. (#1258)
  • If a build uses files CmdStanR can’t check for changes, such as a makefile included from make/local or a header included from the user header, compiling prints a note that changes to them need force_recompile = TRUE. (#1257)
  • pedantic = TRUE now shows the pedantic warnings even when the executable is up to date and isn’t rebuilt. (#1258)
  • Running a model whose executable is out of date, for example because the Stan file was edited after cmdstan_model(), is now an error. Previously the old executable ran. This means the Stan file has to exist whenever the model runs. To run an executable without its Stan file, create the model with cmdstan_model(exe_file = ). (#1255)
  • The new $is_current() method returns FALSE when the executable is out of date, so you can check before running a model, such as one stored in a saved fit. (#1258)
  • A failed compilation now leaves the previous executable in place and usable. Previously the model could end up with no executable, or with one that didn’t match its model methods. (#1235)
  • The compile_standalone and compile_model_methods arguments are gone. Use $expose_functions() and fit$init_model_methods() instead. (#1256, #1245)
  • cmdstan_model(exe_file = ) without a stan_file now gives an error if you also pass cpp_options, stanc_options, include_paths, user_header, force_recompile or pedantic, since there is nothing to compile. The cmdstanr_force_recompile option is ignored for these models. (#1258)
  • cmdstan_model() no longer accepts both stan_file and exe_file, and $exe_file() no longer accepts a path. The executable is named after the Stan file, and the dir argument of cmdstan_model() sets where it goes. (#1253, #1258)
  • $cmdstan_version() now reports the version of CmdStan that compiled the executable, not the version at cmdstan_path(). (#1249)
  • $cpp_options() now returns the options the model was created with, using make variable names and values, so list(stan_threads = TRUE) comes back as list(STAN_THREADS = "true"). Use stan_build_info() to see what the executable reports about its own build. (#1019, #1258)
  • Every cpp_options entry must now be named, as in list(STAN_THREADS = TRUE). For += and the other makefile operators, use cmdstan_make_local(). (#1250)
  • Setting an option to FALSE in cpp_options, as in list(stan_threads = FALSE), now turns it off, even if make/local turns it on. Previously FALSE turned the option on. (#1251)
  • A model compiled with threading now runs when threads_per_chain isn’t set, and asking for more than one thread from a model compiled without threading is now an error. (#1258)
  • Each build setting now goes in one argument, and putting it somewhere else gives an error that names the right one. (#1258)
    • Include paths go in include_paths, not in stanc_options, cpp_options or the STANCFLAGS in make/local.
    • The user header goes in user_header, not in cpp_options. The new $user_header() method returns its path, and $cpp_options() no longer includes it.
    • Other stanc flags go in stanc_options, not in cpp_options as STANCFLAGS.
    • stanc_options no longer accepts flags that have their own argument: include-paths (use include_paths), warn-pedantic (pedantic), allow-undefined (user_header), use-opencl (cpp_options = list(stan_opencl = TRUE)) and name (set from the file name).
  • A flag in stanc_options now overrides the same flag in the STANCFLAGS of make/local. (#1258)
  • stanc_options entries with values, such as list(canonicalize = "deprecations") or list("max-line-length" = 78), now work, including values that contain spaces or quotes. (#1227, #1233, #1263)
  • A stanc error now stops compilation right away and shows the stanc message. (#1227)
  • Relative include paths are now resolved when the model is created, so changing the working directory afterwards no longer breaks #include. $include_paths() returns the absolute paths, and the error for an include path that doesn’t exist shows its absolute path. (#1227, #1229)
  • #include directories with spaces in their paths now work. (#820, #1230)
  • Error messages from a running model now name the model’s Stan file instead of a temporary copy of it. (#1258)
  • $code() and $variables() now always describe the program the model was created from. After $format(overwrite_file = TRUE), call cmdstan_model() again to pick up the rewritten file. (#1258)
  • $expose_functions() and fit$init_model_methods() now work on a model or fit loaded with readRDS(). Previously calling one of the functions or methods failed with a null symbol address error. (#1157, #1158)
  • Compiling, checking syntax and reading a program’s variables no longer leave copies of the program, the generated C++ or the executable in R’s temporary directory. (#1258)

Data and initial values

  • Lists of matrices or vectors, and data frames, can now be passed for int variables, not just real ones. (#817)
  • A data frame column that isn’t numeric, integer, logical or factor is now an error. Previously a character column was silently converted to integer codes in alphabetical order. If that’s what you want, convert the column yourself, e.g. with as.integer(factor(x)). (#1225)
  • Lists of logical vectors or matrices are now converted to integers, like other logical data, instead of giving an error. (#1225)
  • Supplying a factor for a variable not declared as int is now an error. (#1225)
  • Factors are now accepted for length-1 int arrays (e.g. array[1] int x). (#1225)
  • Initial values taken from a Pathfinder fit now use the right weights. Draws CmdStan already resampled with PSIS are no longer reweighted, distinct draws with equal log weights are no longer merged, and a draw that was resampled several times keeps that weight. (#1206, #1207)
  • $pathfinder() now gives each path its own initial values when init supplies one set per path. Previously every path used the first set. (#1206)
  • Initial values taken from a fit now work for a model created with only exe_file. Previously this failed with “‘init’ contains empty lists”. (#1171)
  • A function passed as init is no longer called an extra time to check its output. (#1195)
  • Tuples and complex numbers can now be passed as data and initial values. Pass a tuple as an unnamed list of its elements and a complex number as an R complex value. init = fit now includes tuple parameters. write_stan_json() gains a variables argument, which it needs to tell a tuple from a list of arrays. (#925)

Other new features

  • $sample() can now show a progress bar (using the progressr package) with show_progress_bar = TRUE or options(cmdstanr_progress_bar = TRUE). The option also turns it on when another package such as brms calls CmdStanR for you. The feature is still experimental and may change in the future. (#1138)
  • Reading MCMC and generated quantities draws from CmdStan CSV files now uses less memory. (#1041, #1307)
  • as_cmdstan_fit() now reads draws from the CSV files when they’re first requested (like fits from $sample() do), so looking at a few variables of a large fit takes much less memory. (#1294)
  • read_cmdstan_csv() and as_cmdstan_fit() can now read gzip-compressed CmdStan CSV files ending in .csv.gz and bzip2-compressed files ending in .csv.bz2. CmdStanR now requires data.table 1.18.0 or later. (#1027, #1217)
  • The fitting methods, $save_output_files() and $save_latent_dynamics_files() gain a compress argument for writing gzip- or bzip2-compressed CSV files. The cmdstanr_compress option sets the default for the fitting methods. (#1276, #1311)
  • The new print_stan_file() prints a Stan file, with syntax highlighting when used in a Quarto or R Markdown document. (#1166)
  • The new $cmdstan_defaults() method returns CmdStan’s default argument values for a method ("sample" by default), using CmdStanR’s argument names. (#1167)
  • The new $materialize() method reads a fit’s draws, sampler diagnostics, initial values and profiles from the CSV files into R in one call. (#1181)
  • The new $command() method returns the command line CmdStanR ran for each CmdStan run. (#900)
  • $save_object() gains format = "qs2", which uses the qs2 package and is faster and uses less memory than saveRDS(). (#1125)
  • as_cmdstan_fit() gains a variables argument to read only some of the variables from the CSV files. (#1121)
  • $generate_quantities() gains the show_messages and show_exceptions arguments that the other methods have. (#1142)
  • $loo() now defaults to r_eff = FALSE, skipping the relative efficiency computation, which can be very slow. Set r_eff = TRUE for the previous behavior. (#1091)
  • $log_prob(), $grad_log_prob(), and other model methods are now faster after initialization. (#1274)
  • $init_model_methods() takes less time to set up, and later fits of the same model reuse the compiled methods instead of compiling them again. (#1309)
  • fit$init_model_methods() and $expose_functions() gain a quiet argument that suppresses the messages printed while the methods or functions compile. (#914)
  • install_cmdstan() can now copy the make/local flags of the current installation into the new one, so the new CmdStan is built the same way. Set copy_make_local = TRUE or FALSE, or leave it unset and an interactive session shows the old make/local and asks. (#1267)
  • cmdstan_make_local() no longer adds a flag that’s already in make/local. (#1266)
  • install_cmdstan() now retries a download without the token if GitHub rejects the one in GITHUB_PAT, and warns that the token may be invalid. (#909)
  • The timeout argument of install_cmdstan() now also applies to the download of CmdStan. Previously the download used R’s timeout option (60 seconds by default), which could be too short on a slow connection.
  • Chain IDs in generated file names are now zero-padded to at least two digits, for example 01 instead of 1. (#1244)
  • On Windows, check_cmdstan_toolchain() now finds Rtools through R_TOOLS_SOFT and falls back to PATH, so it works with more R distributions and with future Rtools releases. (#1211)
  • $expose_functions() now gives an error naming the conflicting names when a function or argument name is a reserved word in C++. (#1154)
  • The $expose_functions() documentation now explains how to keep compiled functions and model methods across R sessions with the rcpp.cache.dir option. (#870)
  • The new cmdstanr_save_metric and cmdstanr_save_config options set the defaults of the save_metric and save_cmdstan_config arguments. (#1159)
  • Set the new cmdstanr_spinner option to FALSE to turn off the spinner shown while compiling, checking syntax or installing CmdStan. The spinner is also no longer shown when knitting. (#486)
  • $generate_quantities() now also accepts CmdStanMLE, CmdStanLaplace, and CmdStanPathfinder objects as fitted_params. (#1203)
  • $generate_quantities() now reports per-process execution times with CmdStan 2.39 or newer, and read_cmdstan_csv() returns these times from standalone generated quantities CSV files. (#1168)

Bug fixes

  • $expose_functions() can now expose a function that returns an array of tuples. (#1329)
  • If a fit’s CSV files were in a temporary directory that no longer exists (for example a cached Quarto chunk, or a fit saved with saveRDS() and loaded in a new session), the error now says so and how to avoid it. (#1012)
  • check_cmdstan_toolchain() now waits longer for WSL to respond and says so when it doesn’t, instead of reporting that no WSL distribution is installed. (#1297)
  • $pathfinder(save_single_paths = TRUE) now works. Previously the argument was ignored.
  • $save_metric_files() now gives a clear error when no metric files were written, and saved metric files are no longer deleted when the fit is garbage-collected. (#1021)
  • cmdstan_model() no longer fails when MAKEFLAGS turns on directory printing. (#1163)
  • Quoted values in the STANCFLAGS of make/local now reach stanc as one argument instead of being split on spaces. (#1232)
  • $laplace() with output_basename and no mode no longer writes over the CSV file from the optimization it runs first. That file is now named <output_basename>-mode-01.csv.
  • $laplace() now passes show_messages and show_exceptions to the optimization it runs when mode isn’t supplied.
  • read_cmdstan_csv() now expands ~ in file paths. (#1098)
  • Loading CmdStanR no longer causes parallel::mclapply() to leave zombie processes behind. (#1105)
  • Data with a NULL element now gives an error naming the variable. (#1129)
  • $metadata() now reports the number of chains in num_chains. Previously it was always 1. (#1187)
  • $lp_approx() and $mle() now return numeric vectors whatever the cmdstanr_draws_format option is set to. (#1190)
  • Stan files whose names contain several spaces or quotes now compile. (#1200)
  • $draws() on a Pathfinder fit now orders the diagnostic columns the same way as the other methods. (#1205)
  • An executable that can’t be run, for example one that lost its execute permission or was built for another platform, now gives an error that names the executable and says how to rebuild it. (#1246)
  • A model with no parameters no longer ends every $sample() call with a warning about a NaN E-BFMI.
  • install_cmdstan() now always downloads with R’s libcurl method, so the retry without a rejected GITHUB_PAT also works when options(download.file.method = "curl") is set.
  • On WSL, a failed copy of the executable or the output CSV files now gives an error right away.
  • On Windows, a model now runs with the TBB from the CmdStan installation that built it, even after set_cmdstan_path() switches to another installation. (#1261)
  • Fits with tuple or complex variables now work fully. $metadata() reports their sizes, a tuple’s name selects all of its columns in $draws() and the other methods with a variables argument, and the model methods accept and return them. (#925)
  • With CmdStan in WSL, R can now read the output, diagnostic, profile, config and metric files when output_dir is set, and CmdStan can find data and init files when TMPDIR points at the //wsl$ share. (#1110, #1113)
  • CmdStan processes are now stopped when the R process that started them dies, for example a future worker interrupted from the parent session. Previously they kept running. (#1086)
  • $init_model_methods() and $expose_functions() now work in the same R session as rstan or brms, whichever loads first, when RcppParallel 6.2.0 or later is installed. (#1177, #1270)
  • A function exposed with $expose_functions() that returns a tuple no longer occasionally crashes R or returns a list with wrong elements. (#1001)
  • $expose_functions(global = TRUE) now also puts the functions in the $functions field of the model or fit, so a later $expose_functions() call works. (#1317)
  • cmdstan_make_local(cpp_options = list(stan_threads = FALSE)), and the same argument of install_cmdstan(), now turn threading off. Previously they turned it on. (#605)
  • $init_model_methods() gains a data argument for a fit reloaded in a new session whose data file is gone, and stops before compiling when the data file no longer exists instead of failing with a JSON parsing error afterwards. (#1331)
  • $save_data_file() now leaves the original data file in place. Previously it removed the original, even a file the user had passed as a path. (#1331)

Removed and deprecated

  • CmdStanR now requires R 4.0.0 or later. (#1144)
  • CmdStan versions older than 2.37.0 are no longer supported. To use an older CmdStan, install an older CmdStanR release from GitHub. (#1144)
  • The CMDSTANR_NO_VER_CHECK option and environment variable are deprecated. Use the lowercase cmdstanr_no_ver_check instead.
  • The num_threads argument of $pathfinder() is deprecated. Use threads instead.
  • The save_latent_dynamics argument is now only available in $sample(), $sample_mpi() and $variational(), the methods for which CmdStan writes these files.
  • fit$variable_skeleton() is removed. fit$constrain_variables() now returns the values already structured, so utils::relist() and the skeleton are no longer needed. (#925)
  • data is now the first argument of $init_model_methods(), so a seed passed by position must be named. (#1331)
  • The CMDSTANR_USE_MSYS_TOOLCHAIN environment variable is deprecated and ignored (with a warning). (#1144)
  • Removed deprecated items (replacements in parentheses). (#1061)
    • read_sample_csv() (read_cmdstan_csv())
    • write_stan_tempfile() (write_stan_file())
    • jacobian_adjustment argument to fit$log_prob() and similar methods (jacobian argument)
    • output_samples argument to model$variational() (draws argument)
    • hessian argument to fit$init_model_methods() (the Hessian method is now always compiled)
    • threads (cpp_options = list(stan_threads = TRUE)) and compile_hessian_method (always compiled), formerly arguments to $compile()
    • several arguments to model$sample():
      • cores and num_cores (parallel_chains)
      • num_chains (chains)
      • num_warmup (iter_warmup)
      • num_samples (iter_sampling)
      • validate_csv (diagnostics)
      • save_extra_diagnostics (save_latent_dynamics)
      • max_depth (max_treedepth)
      • stepsize (step_size)

CmdStanR 0.9.0

General Improvements/Changes

  • Added compatibility for RTools45 (#1066)
  • CmdStanR will now use RTools with no additional toolchain updates needed on Windows (CmdStan 2.35+ only; #1065, #1054)
  • Improve error messages when calling sampler_diagnostics() with fixed_param=TRUE
  • Improve numerical stability in calculation of effective sample size during loo method (#1057)
  • Improve numerical stablity with very small log-ratios in calculation of effective sample size during loo method (#1015)
  • Add warning if input data/inits have been coerced to ints (#994)

Bugfixes

  • Don’t require fixed_param for models with zero parameters (only GQs) for CmdStan >= 2.36 (#1046)
  • Improve detection/handling of make (#1036)
  • Fix saving of model objects to network drive (#1038, thanks to @bschneidr)
  • Update usage of untar to fix installation errors (#1034)
  • Respect compilation flags in make/local when exposing functions or model methods (#1003)
  • Fix passing of include paths to CmdStan (#1000)
  • Fix passing of factor data to CmdStan (#999)
  • Fix extraction and passing of array data/parameters as model inits (#993)

Documentation Updates

  • Clarifications to usage of optimize and loo methods (#1060)
  • Add documentation for faster model saving with large models (#1042)
  • Remove mentions of rstan::read_stan_csv due to incompatibility with newer CmdStan outputs (#1018)
  • Document global option cmdstanr_print_line_numbers for printing line numbers (#1017)
  • Change usage of ‘chapter’ to ‘section’ in documentation (#1014)
  • Remove examples of updating removed array syntax as functionality no longer supported in CmdStan (#1008)
  • Change usages of ‘sampling statement’ -> ‘distribution statement’ (#987)

CmdStanR 0.8.1

Minor changes

  • Added CMDSTANR_USE_RTOOLS environment variable to force stock RTools on Windows by @andrjohns in #980
  • Added support for Windows ARM64 by @andrjohns in #990
  • Automatically initialise model methods when called, add inc_warmup argument to $unconstrain_draws() by @andrjohns in #985

Bugfixes

  • Fix errors when using pathfinder object as initial values by @avehtari in #984
  • Fix error with $unconstrain_draws() returning incorrect assumptions in some cases by @andrjohns in #983
  • Fix spurious errors about missing CmdStan config files by @andrjohns in #981
  • Fix linking error when exposing SUNDIALS/KINSOL functions or model methods by @andrjohns in #977
  • Fix long-standing error with OneDrive paths on Windows by @andrjohns in #990

CmdStanR 0.8.0

Major new features

  • Add functionality for passing CmdStanFit objects as initial values by @SteveBronder in #937

Other improvements

  • Add compatibility with CmdStan 2.35 by @andrjohns in #972
  • Add show_messages and show_exceptions arguments to all methods for controlling output by @andrjohns in #897
  • Drop RcppEigen dependency, implement basic Eigen -> C++ interop by @andrjohns in #899
  • Add compatibility with CmdStan 2.34 by @andrjohns in #905 #910
  • Add a format argument to the unconstrain_draws() method to specify draws format of return by @andrjohns in #886
  • Align cmdstanr EBFMI diagnostic threshold with CmdStan by @andrjohns in #892
  • Add global option cmdstanr_print_line_numbers to add line number to model printing by @sbfnk in #967
  • Add new CmdStan arguments save_metric and save_cmdstan_config by @venpopov in #932
  • Add documentation for CmdStanR global options by @jgabry in #951
  • Add documentation for how to obtain structured output similar to rstan::extract() using a combination of cmdstanr and posterior by @jgabry in #955
  • Added coercion generics for CmdStanFit objects by @gowerc in #943
  • psis_resample and calculate_lp arguments added to Pathfinder method by @SteveBronder in #903
  • Documentation and tests for LOO method updated by @jgabry in #923
  • Global option cmdstanr_warn_inits added to disable warnings about partially specified initial values by @jgabry in #913
  • Updates to MCMC output_dir documentation by @jgabry in #929

Bugfixes

  • Fix broken link in OpenCL documentation by @eipi10 in #908
  • Fix a minor typo in the README by @jgabry in #911
  • Make exported RNG functions respect changes to R’s seed by @andrjohns in #973
  • Optimisations for model methods functions by @andrjohns in #960
  • Bugfix for passing function for initial values with Pathfinder method and default num_paths by @andrjohns in #964
  • Continue with compilation if compile_stanalone=TRUE but no functions are found by @jgabry in #956
  • Update tests and CI for compatibility with MacOS ARM64 by @andrjohns in #958
  • Fix handling of inv_metric argument with only 1 parameter by @venpopov in #935
  • Fixes for compatibility with RTools44 by @andrjohns in #952 #959

CmdStanR 0.7.0

Major new features

Other improvements and bug fixes

CmdStanR 0.6.1

  • Store return codes instead of always querying exit status by @jgabry in #798
  • enable jacobian argument for optimization by @jgabry in #799
  • Fix init_model_methods for models with no data by @andrjohns in #801
  • Document a CmdStan-focused way to pre-compile Stan models in R packages by @wlandau in #809
  • Describe how to efficiently save model fit objects by @wlandau in #816
  • fix errors in doc for new methods by @jgabry in #823
  • Give informative error when exposing stan functions with precompiled model by @andrjohns in #831
  • Bugfixes in .stanfunctions, hessian model method, and exposing RNG functions by @andrjohns in #811
  • Fix variable_skeleton() with containers by @andrjohns in #832
  • Improve handling of user header by @martinmodrak in #818
  • change duplicate stdout_file to stderr_file by @jgabry in #834

CmdStanR 0.6.0

Major new features

Other changes

CmdStanR 0.5.3

New features

  • On Windows, users can now install and use CmdStan with WSL (Windows Subsystem for Linux). Set wsl=TRUE in install_cmdstan() to install CmdStan for use with WSL. This can offer significant speedups compared to native Windows execution. (#677, @andrjohns)

Bug fixes

  • In cmdstan_default_path() we now ignore directories inside .cmdstan that don’t start with "cmdstan-". (#651)

  • Fixed Windows issue related to not locating grep.exe or when it is located in a path with spaces. (@weshinsley, #661, #663)

  • Fixed a bug with diagnostic checks when ebfmi is NaN.

  • Fixed a bug that caused issues when using ~ or . in paths supplied to the cmdstanr_write_stan_file_dir global option.

  • Fixed a bug that caused the time() method fail when some of the chains failed to finish succesfully.

CmdStanR 0.5.2

  • Refactored toolchain installation and checks for R 4.x on Windows and added support for Rtools42. (#645)

  • Expanded the use of CMDSTAN environment variable to point to CmdStan installation or directory containing CmdStan installations. (#643)

  • New vignette on how to handle deprecations using the $format() method. (#644)

CmdStanR 0.5.1

  • Temporarily disable format="draws_rvars" in the $draws() method due to a bug. Until this is fixed users can make use of posterior::as_draws_rvars() to convert draws from CmdStanR to the draws_rvars format. (#640)

CmdStanR 0.5.0

Bug fixes

  • Fixed bug that caused stdour/stderr not being read at the end of optimization. (#522)

  • Fixed issue with handling NA as the reported external process status. (#544, @myshkin)

  • Fixed issue with handling models with no parameters and CmdStan 2.27+.

New features

  • Default directory changed to .cmdstan instead of .cmdstanr so that CmdStanPy and CmdStanR can use the same CmdStan installations. Using .cmdstanr will continue to be supported until version 1.0 but install_cmdstan() will now default to .cmdstan and CmdStanR will first look for .cmdstan before falling back on .cmdstanr. (#454)

  • New method diagnose() for CmdstanModel objects exposes CmdStan’s diagnose method for comparing Stan’s gradient computations to gradients computed via finite differences. (#485)

  • New method $variables() for CmdstanModel objects that returns a list of variables in the Stan model, their types and number of dimensions. Does not require the model to be compiled. (#519)

  • New method $format() for auto-formatting and canonicalizing the Stan models. (#625)

  • Added the option to create CmdStanModel from the executable only with the exe_file argument. (#564)

  • Added a convenience argument user_header to $compile() and cmdstan_model() that simplifies the use of an external .hpp file to compile with the model.

  • Added the cmdstanr_force_recompile global option that is used for forcing recompilation of Stan models. (#580)

  • New method $code() for all fitted model objects that returns the Stan code associated with the fitted model. (#575)

  • New method $diagnostic_summary() for CmdStanMCMC objects that summarizes the sampler diagnostics (divergences, treedepth, ebfmi) and can regenerate the related warning messages. (#205)

  • New diagnostics argument for the $sample() method to specify which diagnostics are checked after sampling. Replaces validate_csv argument. (#205)

  • Added E-BFMI checks that run automatically post sampling. (#500, @jsocolar)

  • New methods for posterior::as_draws() for CmdStanR fitted model objects. These are just wrappers around the $draws() method provided for convenience. (#532)

  • write_stan_file() now choose file names deterministically based on the code so that models do not get unnecessarily recompiled when calling the function multiple times with the same code. (#495, @martinmodrak)

  • The dir argument for write_stan_file() can now be set with a global option. (#537)

  • write_stan_json() now handles data of class "table". Tables are converted to vector, matrix, or array depending on the dimensions of the table. (#528)

  • Improved processing of named lists supplied to the data argument to JSON data files: checking whether the list includes all required elements/Stan variables; improved differentiating arrays/vectors of length 1 and scalars when generating JSON data files; generating floating point numbers with decimal points to fix issue with parsing large numbers. (#538)

  • install_cmdstan() now automatically installs the Linux ARM CmdStan when Linux distributions running on ARM CPUs are detected. (#531)

  • New function as_mcmc.list() for converting CmdStanMCMC objects to mcmc.list objects from the coda package. (#584, @MatsuuraKentaro)

CmdStanR 0.4.0

Bug fixes

  • Fixed issue with retrieving draws with models with spaces in their names. (#453)

  • Fixed bug with spaces in path to the temporary folder on Windows. (#460)

  • Fixed issue with not reporting model executable name clashing with folder name. (#461)

New features

  • New function as_cmdstan_fit() that creates CmdStanMCMC/MLE/VB objects directly from CmdStan CSV files. (#412)

  • read_cmdstan_csv() now also returns chain run times for MCMC sampling CSV files. (#414)

  • Faster CSV reading for multiple chains. (#419)

  • New $profiles() method for fitted model objects accesses profiling information from R if profiling used in the Stan program. Support for profiling Stan programs requires CmdStan >= 2.26. (#434)

  • New vignette on profiling Stan programs. (#435)

  • New vignette on running Stan on the GPU with OpenCL. OpenCL device ids can now also be specified at runtime. (#439)

  • New check for invalid parameter names when supplying init values. (#452, @mike-lawrence)

  • Suppressing compilation messages when not in interactive mode. (#462, @wlandau)

  • New error_on_NA argument for cmdstan_version() to optionally return NULL (instead of erroring) if the CmdStan path is not found (#467, @wlandau).

  • Global option cmdstanr_max_rows can be set as an alternative to specifying max_rows argument to the $print() method. (#470)

  • New output_basename argument for the model fitting methods. Can be used in conjunction with output_dir to get completely predictable output CSV file paths. (#471)

  • New format argument for $draws(), $sampler_diagnostics(), read_cmdstan_csv(), and as_cmdstan_fit(). This controls the format of the draws returned or stored in the object. Changing the format can improve speed and memory usage for large models. (#482)

CmdStanR 0.3.0

Bug fixes

  • Fixed reading inverse mass matrix with values written in scientific format in the CSV. (#394)

  • Fixed error caused by an empty data list. Previously if a model didn’t require data then data had to either be NULL or be a non-empty list, but now list() is allowed. (#403)

New features

  • Added $sample_mpi() for MCMC sampling with MPI. (#350)

  • Added informative messages on compile errors caused by precompiled headers (PCH). (#384)

  • Added the cmdstanr_verbose option for verbose mode. Intended for troubleshooting, debugging and development. See end of How does CmdStanR work? vignette for details. (#392)

  • New $loo() method for CmdStanMCMC objects. Requires computing pointwise log-likelihood in Stan program. (#366)

  • The fitted_params argument to the $generate_quantities() method now also accepts CmdStanVB, posterior::draws_array, and posterior::draws_matrix objects. (#390)

  • The $optimize() method now supports all of CmdStan’s tolerance-related arguments for (L)BFGS. (#398)

  • The documentation for the R6 methods now uses @param, which allows package developers to import the CmdStanR documentation using roxygen2’s @inheritParams. (#408)

CmdStanR 0.2.2

Bug fixes

  • Fixed bug with reading Stan CSV when grep used coloring by default (#364,#371)

  • Depend on posterior v0.1.3 to avoid a potential error in $summary(). (#383)

New features

  • Added support for native execution on the macOS with the M1 ARM-based CPU. (#375)

  • Added threading support via threads argument for $optimize() and $variational() (was already available via threads_per_chain for $sample()). (#369)

CmdStanR 0.2.1

Bug fixes

New features

  • compile() and check_syntax() methods gain argument pedantic for turning on pedantic mode, which warns about issues with the model beyond syntax errors. (#361)

CmdStanR 0.2.0

Bug fixes

  • Fix potential indexing error if using read_cmdstan_csv() with CSV files created by CmdStan without CmdStanR. (#291, #292, @johnlees)

  • Fix error when returning draws or sampler diagnostics for a fit with only warmup and no samples. (#288, #293)

  • Fix trailing slashes issue for dir in cmdstan_model() and output_dir in fitting methods. (#281, #294)

  • Fix dimensions error when processing a list of matrices passed in as data. (#296, #302)

  • Fix reporting of time after using fixed_param method. (#303, #307)

  • With refresh = 0, no output other than error messages is printed with $optimize() and $variational(). (#324)

  • Fix issue where names of generated files could clash. (#326, #328)

  • Fix missing include_paths in $syntax_check(). (#335, @mike-lawrence)

New features

  • CSV reading is now faster by using data.table::fread(). (#318)

  • install_cmdstan() gains argument version for specifying which version of CmdStan to install. (#300, #308)

  • New function check_cmdstan_toolchain() that checks if the appropriate toolchains are available. (#289)

  • $sample() method for CmdStanModel objects gains argument chain_ids for specifying custom chain IDs. (#319)

  • Added support for the sig_figs argument in CmdStan versions 2.25 and above. (#327)

  • Added checks if the user has the necessary permissions in the RTools and temporary folders. (#343)

CmdStanR 0.1.3

  • New $check_syntax() method for CmdStanModel objects. (#276, #277)

CmdStanR 0.1.2

  • User is notified by message at load time if a new release of CmdStan is available. (#265, #273)

  • write_stan_file() replaces write_stan_tempfile(), which is now deprecated. With the addition of the dir argument, the file written is not necessarily temporary. (#267, #272)

CmdStanR 0.1.1

CmdStanR 0.1.0

  • Beta release