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.

Each entry says what changed and what it replaced, so you (or your coding agent) can update existing code from them.

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() likewise work from a Stan file alone. (#1256)
  • cmdstan_model() recompiles whenever anything it tracks has changed: the Stan program and its #includes, the user header, cpp_options, stanc_options, make/local or the CmdStan installation. In an interactive session, or with quiet = FALSE, a message says which. Changes it cannot see, such as a new C++ compiler, need force_recompile = TRUE. The vignette “How does CmdStanR work?” lists them. (#1255, #1237, #1019, #1308)
  • Each executable now comes with a build record, a hidden JSON file beside it that says how it was built: bernoulli gets .bernoulli.cmdstanr.json and bernoulli.exe gets .bernoulli.exe.cmdstanr.json. Add .*.cmdstanr.json to .gitignore wherever the executable is already ignored. If the record is missing, cmdstan_model() recompiles. The vignette “How does CmdStanR work?” has the details. (#1238)
  • The new stan_build_info() shows how an executable was built: the options it was asked for, the files it read, the CmdStan installation that built it, and what the executable itself reports. Without a usable record it says why and shows only what the executable reports. $build_info() does the same for a model object. (#1258)
  • When a build depends on files CmdStanR cannot track, such as a makefile included from make/local or a header included from the user header, the build prints a one-line note saying that changes to them need force_recompile = TRUE. (#1257)
  • pedantic = TRUE now runs the pedantic check whether or not the executable is rebuilt. Previously an up-to-date executable skipped the build and the check with it. (#1258)
  • Running a model whose executable is out of date is now an error that points at cmdstan_model(). Previously the stale executable ran as if it were current. Because the check reads the Stan file, a model created from one needs that file whenever it runs. To run an executable without its Stan file, create the model with cmdstan_model(exe_file = ). (#1255)
  • The new $is_current() method runs that same check and returns TRUE or FALSE instead of raising that error, so a package that keeps a CmdStanModel inside a saved fit can decide whether to rebuild before running it. (#1258)
  • A failed compilation leaves the previous executable in place. Previously a failure at the C++ stage could leave the old executable paired with model methods generated from the new program, and a failed installation could leave the model with no executable at all. (#1235)
  • The compile_standalone and compile_model_methods arguments are gone. Previously both did nothing when the executable was already up to date. Call $expose_functions() and fit$init_model_methods() instead, which work whether the executable was just compiled or reused. (#1256, #1245)
  • cmdstan_model(exe_file = ) with no stan_file now rejects cpp_options, stanc_options, include_paths, user_header, force_recompile and pedantic: with no Stan file there is nothing to compile, so the executable is used as it is. The cmdstanr_force_recompile option is ignored for such a model. (#1258)
  • cmdstan_model() no longer accepts stan_file and exe_file together. Use dir to choose the directory the executable is built in. The executable is named after the Stan file. Previously exe_file beside stan_file set the path the executable was built at. (#1258)
  • $exe_file() no longer accepts a path. Where the executable goes is set when the model is created, with dir. (#1253)
  • $cmdstan_version() now reports the version of CmdStan that compiled the executable, not the version at cmdstan_path(). (#1249)
  • $cpp_options() now returns exactly the options the model was created with, spelled as make variables: list(stan_threads = TRUE) comes back as STAN_THREADS, with the value "true" (the string make received, so FALSE comes back as ""). What the executable reports about its own build has moved to stan_build_info(). (#1019, #1258)
  • Every cpp_options entry must now be named, with a make variable name. An unnamed entry gets an error saying where it belongs: list(NAME = value) for a plain assignment, cmdstan_make_local() for += and the other makefile operators. Previously unnamed entries reached make but nothing else saw them. (#1250)
  • cpp_options = list(stan_threads = FALSE) now turns threading off, even when make/local turns it on, and the same holds for FALSE on any option. Previously FALSE was passed through as a value, which enabled the option. (#1251)
  • Running a model compiled with threading no longer errors when threads_per_chain is not set. Asking for more than one thread from a model compiled without threading is now an error. The thread count also no longer lingers in the session’s environment after the call. (#1258)
  • Each build setting now has one place to go, and putting it anywhere else is an error that points at the right argument. (#1258)
    • Include paths go in include_paths, not in stanc_options, in cpp_options as STANCFLAGS or in make/local’s STANCFLAGS. Previously a model with include paths in stanc_options compiled and then failed on $sample().
    • 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 rejects the flags CmdStanR’s own arguments set: include-paths (use include_paths), warn-pedantic (pedantic), allow-undefined (user_header), use-opencl (cpp_options = list(stan_opencl = TRUE)) and name, which comes from the file name.
  • A flag in stanc_options now overrides the same flag in make/local’s STANCFLAGS. (#1258)
  • Named stanc_options values such as list(canonicalize = "deprecations") and numeric ones such as list("max-line-length" = 78) now work. Previously the named values reached stanc shell-quoted, which it rejected, and the numeric ones were dropped. A value holding a space, a quote or a $ now also reaches stanc intact through make. (#1227, #1233, #1263)
  • A stanc error now stops the build immediately and shows stanc’s message. Previously it surfaced several steps later. (#1227)
  • An include path that does not exist is now reported by its absolute path. (#1227)
  • $include_paths() now returns absolute paths, resolved when the model is created. Previously a relative include path was resolved on every stanc call, so changing the working directory could point #include at the wrong directory. (#1229)
  • #include directories with spaces in their paths now work. (#820, #1230)
  • Error messages from a running model now name the Stan file the model was created from. Previously they named a temporary copy of the program that no longer existed. (#1258)
  • $code() and $variables() now describe the program the model was created from, even after $format(overwrite_file = TRUE) rewrites the file. Create the model again to pick up the change. Previously $format(overwrite_file = TRUE) replaced what $code() returned. (#1258)
  • $expose_functions() and fit$init_model_methods() now work on a model or fit loaded with readRDS(), compiling the functions and methods again. Previously they claimed to be compiled already and calling one failed with a null symbol address. (#1157, #1158)
  • Compiling no longer leaves a copy of the Stan program, the generated C++ and a second copy of the executable in the session’s temporary directory, and checking syntax or reading a program’s variables no longer leaves stanc’s output there. (#1258)

Data and initial values

  • Lists of matrices/vectors and data frames can now be supplied for variables declared as int in the Stan program. Previously these worked only for real variables and errored for int ones. (#817)
  • Data frame columns that are not numeric, integer, logical, or factor are now an error. Previously data.matrix() silently coerced them, so a character column reached Stan as alphabetically ordered integer codes. Convert the column explicitly, e.g. with as.integer(), if integer codes are what you want. (#1225)
  • Lists of logical vectors/matrices are now converted to integers like logical variables are, instead of erroring. (#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), which previously errored. (#1225)
  • Pathfinder fits used as initial values now use uniform weights when CmdStan already PSIS-resampled their draws, avoiding a second application of importance weights. (#1206)
  • Pathfinder fits used as initial values now correctly treat draws with different initialization parameter values as distinct even when their log weights are equal, and collapse duplicate resampled draws while retaining their selection frequency. (#1207)
  • pathfinder() now passes separately supplied initial values to every path instead of using only the first path’s initial values. (#1206)
  • Initial values taken from a fitted model object now work for a model created with only exe_file. Previously the error was “‘init’ contains empty lists”. (#1171)
  • Functions supplied as init are no longer called an extra time to validate them. Validation occurs in the same pass. (#1195)
  • Tuple and complex values can now be passed as data and initial values, a tuple as an unnamed list of its elements and a complex number as an R complex value, and init = fit now carries tuple parameters over. write_stan_json() gains a variables argument for the declarations it needs to tell a tuple from a list of arrays. (#925)

Other new features

  • $sample() can show a progress bar through 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. (#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). Reading a large fit to look at a few variables 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)
  • All fitting methods plus $save_output_files() and $save_latent_dynamics_files() gain a compress argument that writes the CSV files gzip- or bzip2-compressed, with the cmdstanr_compress option setting 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, under the corresponding CmdStanR 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 saves with the qs2 package instead of saveRDS() and is faster and uses less memory. (#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() now offers to copy the make/local flags of the current installation into the new one before building it, so the new CmdStan is built with the same flags. In an interactive session it shows the previous make/local and asks if the copy_make_local argument isn’t set to TRUE or FALSE. (#1267)
  • cmdstan_make_local() now skips flags that are already in make/local. Previously copying the flags of a previous installation after every upgrade added the same lines again each time. (#1266)
  • install_cmdstan() now retries a download without the token if GitHub rejects the one in GITHUB_PAT, and warns the token may not be valid. (#909)
  • Chain IDs in generated file names are now zero-padded to at least two digits, for example 01 instead of 1. (#1244)
  • When using CmdStan through WSL, paths for output, diagnostic, profile, config, and metric files now remain accessible to Windows R when an explicit output directory is supplied. (#1110; related: #1113)
  • check_cmdstan_toolchain() now locates Windows toolchains using R_TOOLS_SOFT and falls back to PATH, improving support for alternate R distributions and future Rtools releases. (#1211)
  • Exposing functions using names that are reserved keywords now throws an informative error message. (#1154)
  • Compiled Stan functions and model methods can be kept across R sessions by setting the rcpp.cache.dir option, as the $expose_functions() documentation now explains. (#870)
  • save_cmdstan_config and save_metric default to FALSE but can be set to TRUE for an entire R session via new global options. (#1159)
  • The compilation spinner can now be disabled for an entire R session by setting the new cmdstanr_spinner global option to FALSE. The spinner shown while installing or rebuilding CmdStan and while checking syntax also respects this option, and is no longer shown when knitting. (#486)
  • $generate_quantities() now also accepts CmdStanMLE, CmdStanLaplace, and CmdStanPathfinder fitted model 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

  • A fit whose CSV files were written to the temporary directory and are gone (a cached Quarto chunk, a fit saved with saveRDS()) now says so and how to avoid it, instead of failing with “File does not exist”. (#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() now respects save_single_paths = TRUE instead of always passing 0 to CmdStan.
  • The save_latent_dynamics argument is now limited to $sample(), $sample_mpi(), and $variational(), matching the CmdStan algorithms that support diagnostic CSV output.
  • save_metric_files() now gives an informative error when metric files were not created and keeps saved metric files after the fitted model is garbage-collected. (#1021)
  • cmdstan_model() no longer fails when MAKEFLAGS turns on directory printing. (#1163)
  • Quoted values in make/local’s STANCFLAGS now reach stanc as one argument. Previously they were split on whitespace. (#1232)
  • laplace() no longer overwrites the internally generated optimizer CSV when mode = NULL and output_basename is supplied. The internally generated optimizer CSV now uses the filename <output_basename>-mode-01.csv.
  • $laplace() now passes show_messages and show_exceptions to the optimizer run it starts when mode is not supplied. Previously that run printed its output regardless.
  • read_cmdstan_csv() now expands ~ in file paths. (#1098)
  • Loading CmdStanR no longer causes parallel::mclapply() to leave zombie processes behind. (#1105)
  • Supplying data with a NULL element now gives an informative error. (#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)
  • A Stan file name with several spaces, or quotes, now gives a valid model name for stanc. Previously only the first space was replaced and the quotes ended up in the generated C++. (#1200)
  • $draws() on a pathfinder fit now orders the diagnostic columns the same way as the other methods. (#1205)
  • An executable that cannot be run, for example one that lost its execute bit or was built for another platform, now gives an error naming the executable and saying how to rebuild it. Previously the fitting methods and $cmdstan_defaults() surfaced a raw processx error. (#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. Previously, with the option set to "curl", a bad GITHUB_PAT left a GitHub error page in place of the download instead of triggering a retry without the token.
  • On WSL a failed copy of the executable or the output CSV files now gives an error at the copy instead of later when the run or read can’t find the file.
  • On Windows a model executable is now launched with the TBB it was built against. Previously the selected CmdStan installation’s TBB was used, which was wrong once set_cmdstan_path() had selected a different one. (#1261)
  • Fits with tuple or complex variables now work throughout: $metadata() reports their names and sizes, a tuple’s name selects all of its columns in $draws() and the other methods with a variables argument, and the model methods accept and return them. Previously reading such a fit warned about NAs in the variable sizes. (#925)
  • When using CmdStan through WSL, a data or init file on the WSL filesystem is now found when R spells its path with backslashes, as it does for temporary files when TMPDIR points at the //wsl$ share. (#1113)
  • CmdStan processes are now killed when the R process that started them dies without running its cleanup, for example a future worker interrupted from the parent session. Previously they kept running as orphans. (#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 keeps the functions in the $functions field of the model or fit. Previously it left that field empty while marking the functions as compiled, so a later $expose_functions() call exposed nothing. (#1317)
  • cmdstan_make_local(cpp_options = list(stan_threads = FALSE)), and the same argument of install_cmdstan(), now turn the option off by writing STAN_THREADS= to make/local. Previously they wrote STAN_THREADS=false, which CmdStan reads as on. (#605)

Removed and deprecated

  • Minimum R version increased to 4.0.0. (#1144)
  • CmdStan versions older than 2.37.0 are no longer supported. To use an older CmdStan version install an older CmdStanR release from GitHub. (#1144)
  • The CMDSTANR_NO_VER_CHECK R option and environment variable are deprecated as of CmdStanR 1.0.0; use the lowercase cmdstanr_no_ver_check forms instead.
  • pathfinder() now uses the threads argument (num_threads is deprecated), to be consistent with other methods.
  • 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)
  • Removed legacy Windows toolchain paths for older CmdStan releases. (#1144)
  • CMDSTANR_USE_MSYS_TOOLCHAIN is now 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() (hessian method always compiled now)
    • 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