Getting Started#

Requirement: C++ toolchain#

Stan requires a C++ tool chain consisting of

  • A C++14 compiler. On Windows, MSCV is not supported, so something like MinGW GCC is required.

  • The Gnu make utility. On Windows, make or mingw32-make can be used

Here are complete instructions by platform for installing both, from the CmdStan installation instructions.

Downloading BridgeStan#

Note

The Julia, Python, and R interfaces will download the source for you the first time you compile a model, and the Rust interface has an optional feature to download the source upon request. This section is optional for users primarily interested in those interfaces.

Installing BridgeStan is as simple as ensuring that the above requirements are installed and then downloading the source repository. All of the following ways of downloading BridgeStan will additionally download the Stan and Stan Math libraries for you, and no additional dependencies are required to be installed separately for the C++ source code.

Downloading a released archive#

Downloads of a complete copy of the source code and interfaces are available on our GitHub releases page.

To use these, simply download the file associated with the version you wish to use, and unzip its contents into the folder you would like BridgeStan to be in.

Installing the latest version with git#

If you have git installed, you may download BridgeStan by navigating to the folder you’d like BridgeStan to be in and running

git clone --recurse-submodules --shallow-submodules --depth=1 https://github.com/roualdes/bridgestan.git

If you clone without the --recurse-submodules argument, you can download the required submodules with make stan-update. The arguments --shallow-submodules and --depth=1 are to reduce the size of the download, but are not required.

Testing the Installation#

After this, BridgeStan is installed. You can test a basic compilation by opening a terminal in your BridgeStan folder and running

make test_models/multi/multi_model.so

This will compile the file test_models/multi/multi.stan into a shared library object for use with BridgeStan. This will require internet access the first time you run it in order to download the appropriate Stan compiler for your platform into <bridgestan-dir>/bin/stanc[.exe]

Installing an Interface#

To see instructions for installing the BridgeStan client package in your language of choice, see the Language Interfaces page.

Optional: Customizing BridgeStan#

BridgeStan has many compiler flags and options set by default. Many of these defaults are the same as those used by the CmdStan interface to Stan. You can override the defaults or add new flags on the command line when invoking make, or make them persistent by creating or editing the file <bridgestan dir>/make/local.

For example, setting the contents of make/local to the following includes compiler flags for optimization level and architecture.

# By default we use -O3, this sets a less aggressive C++ optimization level
O=2
# Adding other arbitrary C++ compiler flags
CXXFLAGS+= -march=native

Flags for stanc3 can also be set here

# pedantic mode and level 1 optimization
STANCFLAGS+= --warn-pedantic --O1

Enabling Parallel Calls of Stan Programs#

In order for Python or Julia to be able to call a single Stan model concurrently from multiple threads or for a Stan model to execute its own code in parallel, the following flag must be set in make/local or on the command line.

# Enable threading
STAN_THREADS=true

Note that this flag changes a lot of the internals of the Stan library and as such, all models used in the same process should have the same setting. Mixing models which have STAN_THREADS enabled with those that do not will most likely lead to segmentation faults or other crashes.

Additional flags, such as those for MPI and OpenCL, are covered in the CmdStan User’s Guide page on Parallelization.

Autodiff Hessian calculations#

By default, Hessians in BridgeStan are calculated using central finite differences. This is because not all Stan models support the nested autodiff required for Hessians to be computed directly, particularly models which use implicit functions like the algebra_solver or ODE integrators.

If your Stan model does not use these features, you can enable autodiff Hessians by setting the compile-time flag BRIDGESTAN_AD_HESSIAN=true in the invocation to make. This can be set in make/local if you wish to use it by default.

This value is reported by the model_info function if you would like to check at run time whether Hessians are computed with nested autodiff or with finite differences. Similar to STAN_THREADS, it is not advised to mix models which use autodiff Hessians with those that do not in the same program.

Autodiff Hessians may be faster than finite differences depending on your model, and will generally be more numerically stable.

Constraint tolerances#

The param_unconstrain family of functions check their inputs to ensure that they are in the support of the unconstraining transform. For example, if the model has a simplex parameter, it will verify that all of the elements sum to 1.0.

When unconstraining outputs from e.g. CmdStan, it is not uncommon that these constraints are violated by a small amount due to numerical error. If this amount is larger than 1e-8, the function will throw an error.

This tolerance is set in the Stan Math library, but can be overridden by defining the STAN_MATH_CONSTRAINT_TOLERANCE during compilation. One way to do this in BridgeStan is to set CPPFLAGS in make/local:

CPPFLAGS+=-DSTAN_MATH_CONSTRAINT_TOLERANCE=1e-5

Using External C++ Code#

BridgeStan supports the same capability to plug in external C++ code as CmdStan.

Namely, you can declare a function in your Stan model and then define it in a separate C++ file. This requires passing the --allow-undefined flag to the Stan compiler when building your model. The USER_HEADER variable must point to the C++ file containing the function definition. By default, this will be the file user_header.hpp in the same directory as the Stan model.

For a more complete example, consult the CmdStan documentation.

Using Older Stan Versions#

If you wish to use BridgeStan for an older released version, all you need to do is

  1. Set STANC3_VERSION in make/local to your desired version, e.g. v2.26.0

  2. Go into the stan submodule and run git checkout release/VERSION, e.g. release/v2.26.0

  3. Also in the stan submodule, run make math-update

  4. In the top level BridgeStan directory, run make clean

To return to the version of Stan currently used by BridgeStan, you can run make stan-update from the top level directory and remove STANC3_VERSION from your make/local file, before running make clean again.

Using Pre-Existing Stan Installations#

If you wish to use BridgeStan with a pre-existing download of the Stan repository, or with a custom fork or branch, you can set the STAN (and, optionally, MATH) variables to the path to your existing copy in calls to make, or more permanently by setting them in a make/local file as described above.

The easiest way to use a custom stanc3 is to place the built executable at bin/stanc[.exe].