The new Stan compiler features a TensorFlow Probability backend, transpiling Stan programs to python code.
stan2tfp is a lightweight interface wrapper around this functionality, allowing users to:
- call the compiler (emitting TFP code)
- run the code (creating a model object in the current namespace)
- sample the model (using TFP's NUTS)
... without leaving the notebook or their favorite IDE.
The new compiler and the TFP backend are under active development. Currently only a small subset of Stan's functionality is supported. For a list of supported distributions, see here.
stan2tfp is a pure-Python package which can be installed from PyPI
pip install stan2tfp
This will also install TensorFlow and TensorFlow Probability (both in nightly version; needed for XLA compilation of the model).
stan2tfp provides the function download_stan2tfp_compiler which downloads a pre-compiled binary of the compiler. By default it installs the latest version into the same directory as the package itself. A different path or a different version of the compiler can be specified.
For a simple end-to-end example of using stan2tfp, see here.