.. py:currentmodule:: cmdstanpy
For full changes, see the Releases page on GitHub.
- Fixed issues that arose when running a model with no parameters using CmdStan 2.36.0+
Reminder: The next non-bugfix release of CmdStanPy will be version 2.0, which will remove all existing deprecations.
- Fixed a bug in from_csv which prevented reading files created by CmdStan 2.35.0+
Reminder: The next non-bugfix release of CmdStanPy will be version 2.0, which will remove all existing deprecations.
- Updated the logic around reading Stan CSV files to support CmdStan 2.35.0+
- Fixed an issue where the
profile_filesmember of the RunSet object was not correct when running multiple chains in the same process.
Reminder: The next non-bugfix release of CmdStanPy will be version 2.0, which will remove all existing deprecations.
- Updated Community page to link to the
bibatpackage. - Moved CmdStanPy's metadata to exclusively use the
pyproject.tomlfile. - Fixed an issue where the deprecation of the
compile=Falseargument to :class:`CmdStanModel` would make it impossible to use the canonicalizer to update old models. The new function :func:`cmdstanpy.format_stan_file` supports this use case. - Fixed a bug preventing multiple inits from being used with :meth:`CmdStanModel.pathfinder`.
- Added a helper argument
num_threadsto :meth:`CmdStanModel.pathfinder`.
Reminder: The next non-bugfix release of CmdStanPy will be version 2.0, which will remove all existing deprecations.
- Switched from :class:`numpy.random.RandomState` to :func:`~numpy.random.default_rng`.
- Fixed minor doc typos.
- Stan 2.34: Fixed parsing of unit_e output files.
- Stan 2.34: Exposed new Pathfinder arguments.
- Allow the
varsargument to :meth:`CmdStanMCMC.draws_pd` to filter thechain__,iter__, anddraws__columns. - Fixed a bug involving the interaction of the
adapt_engagedanditer_warmuparguments to :meth:`CmdStanMCMC.sample`.
Reminder: The next non-bugfix release of CmdStanPy will be version 2.0, which will remove all existing deprecations.
- New functionality
- The Pathfinder algorithm (available in CmdStan 2.33+) is now availble as :meth:`CmdStanModel.pathfinder`.
- Laplace approximations (available in CmdStan 2.32+) are now available as :meth:`CmdStanModel.laplace_sample`.
- The :meth:`CmdStanModel.optimize` method now supports the
jacobianboolean argument to enable change-of-variables adjustments. When enabled, the Maximum a posteriori estimate (MAP) is returned, rather than the MLE. - The :func:`cmdstanpy.install_cmdstan` function and script can install development versions of CmdStan using the
git:prefix in the version.
- Deprecations
The next non-bugfix release of CmdStanPy will be version 2.0, which will remove all existing deprecations. Additional deprecations in this version:
- :class:`CmdStanModel` will require that it has a compiled executable after construction. The
compileargument is deprecated, (the ability to force recompilation is available under the argumentforce_compile), and thecompile()method is deprecated. If you wish to compile Stan files independent of constructing a model, use :func:`cmdstanpy.compile_stan_file`. - :meth:`CmdStanMLE.stan_variable` will begin returning a :class:`np.ndarray` in all cases, as opposed to the current behavior where sometimes a float is returned.
- :meth:`CmdStanVB.stan_variables` will return the _draws_ from the approximate posterior, rather than the optimized mean.
A new argument,
mean, can be set to True to return the mean instead. Additionally, a :class:`np.ndarray` will be returned in all cases starting in the next version. - :meth:`CmdStanModel.variational` argument
output_sampleswill has been renamed todraws.
- :class:`CmdStanModel` will require that it has a compiled executable after construction. The
- Other changes
- A list of dictionaries is now allowed as the
initsargument to :meth:`CmdStanModel.sample`. - :func:`cmdstanpy.install_cmdstan` correctly fetches the CmdStan version for ppc64el machines.
- The documentation on how to use external C++ code was updated.
- Various other bug fixes.
- A list of dictionaries is now allowed as the
Note
The minimum supported version for CmdStanPy is now Python 3.8.
- New functionality
- :meth:`CmdStanModel.generate_quantities` can now accept samples from optimization and variational inference.
The argument
mcmc_samplehas been renamedprevious_fitto reflect this; the former name is still accepted but deprecated. - :meth:`CmdStanModel.log_prob` is able to return the log probability and its gradient with respect to a set of parameters. Note that this is not an efficient way of calculating this in general and should be reserved for debugging and model development.
- :meth:`CmdStanModel.generate_quantities` can now accept samples from optimization and variational inference.
The argument
- Other changes
- Improved some of the type hints in the package.
- Ensure draws are serialized if a fit object is pickled.
- :meth:`~CmdStanModel.src_info` now raises an error if the command fails, rather than returning
{}. - CmdStanPy has transitioned all unit tests from the
unittestlibrary to usepytest.
install_cmdstannow downloads the correct CmdStan for non-x86 Linux machines.- Improved reporting of errors during :meth:`~CmdStanModel.compile`.
- Fixed some edge cases in mixing arguments of the :meth:`~CmdStanModel.optimize` function.
- Fixed how
NaNand infinite numbers were serialized to JSON. - Removed dependency on
ujson. For now, all JSON serialization is done with the Python standard library. - Added a
timeoutparameter to all model methods which can be used to terminate the CmdStan process after the specified time. - A model will now properly recompile if one of the #include-d files changed since it was last built.
- Fixed an issue where complex number containers in Stan program outputs were not being read in properly by CmdStanPy. The output would have the correct shape, but the values would be mixed up.
- Fixed a build error in the documentation
- Improved messages when model fails to compile due to C++ errors.
- Fixed a typo in :func:`cmdstanpy.show_versions()`
- Reorganized and updated the documentation
- Reorganized a lot of internal code
- Cleaned up the output of :meth:`CmdStanMCMC.draws_pd`
- Cleaned up the output of :meth:`CmdStanMCMC.summary`
- Removed the logging which occurred when Python exited with cmdstanpy imported.
- Fix an issue with :func:`cmdstanpy.install_cmdstan()` where the installation would report that it had failed even when it had not.
Fix an issue where Stan fit objects were not
pickle-able when they previously were.Warning
We still do not recommend pickling cmdstanpy objects, but rather using functions :meth:`~CmdStanMCMC.save_csvfiles` and :func:`~cmdstanpy.from_csv`.
- CmdStanPy can now format (and canonicalize) your Stan files with :meth:`CmdStanModel.format()`
- Stan variables can now be accessed from fit objects using the . syntax when no naming conflicts occur. For example, previous code
fit.stan_variable("my_cool_variable")can now be writtenfit.my_cool_variable - CmdStanPy is more robust to running in threaded environments and tries harder to not overwrite its own output files
- The
install_cmdstanscript can now be run in interactive mode using--interactive/-i - CmdStanPy now computes some diagnostics after running HMC and will warn you about post-warmup divergences and treedepth exceptions
- Runtime exceptions in the
generated quantitiesblock should be recognized better now. - The default level of precision used by :meth:`CmdStanMCMC.summary()` is now 6, as it is when
stansummaryis used from the command line. - Various documentation improvements
- Support new optimizations in CmdStan 2.29
- Support complex numbers as both inputs and outputs of Stan programs
- Sped up assembling output by only reading draws at most once
- Fixed an issue where a command failing could change your working directory
- Improve error messages in some cases
- CmdStanPy no longer changes the global root logging level
Note
The minimum supported version for CmdStanPy is now Python 3.7.
- Initial release