Yet Another Bayesian Framework
- Free software: MIT license
- Documentation: https://yabf.readthedocs.io.
Why another Bayesian Framework? There are a relative host of Bayesian codes in Python, including major players such as PyMC3, emcee and stan, as well as a seemingly never-ending set of scientific-field-specific codes (eg. cosmology has CosmoMC, CosmoHammer, MontePython, cobaya...).
yabf
was written because the author found that all the frameowrks they tried
were either too lean or too involved. yabf
tries to find the happy medium.
It won't be the right tool for everyone, but it might be the right tool for you.
yabf
is designed to support "black box" likelihoods, by which we mean those
that don't necessarily have analytic derivatives. This separates it from codes
such as PyMC3 and stan, and limits its use to samplers that do not require
that information. This is more often the case in scientific applications, where
likelihoods can in principle depend on some enormous black-box simulation code.
Thus, in this regard it is more like emcee or polychord.
On the other hand, yabf
is not another MCMC sampler. Apart from the
limitations concerning likelihood derivatives, it is sampler-agnostic. It is
rather a specification of a format, and an implementation of that specification.
That is, it specifies that likelihoods should have certain properties (like
parameters), and gives tools that enable that. Or as another example, it
specifies that samplers should contain certain attributes pre- and post-sampling.
In this regard, yabf
is more like PyMC3 or stan, and unlike emcee or
polychord.
yabf
is perhaps most similar to codes such as CosmoHammer or cobaya,
which provide an interface for creating (cosmological) likelihoods which can
then be sampled by somie specified sampler. However, yabf
is different in
that it is intended to be field-agnostic, and entirely general. In addition,
I found that these codes didn't quite satisfy my criteria for ease-of-use
and extensibility.
I hope that yabf
provides these. Here are a few of its features:
- Deisgn is both DRY/modular and easy-to-use: while components of the model can be separately defined (to make it DRY), they don't need to be combined into a rigid structure in order to perform most calculations. This makes it easy to evaluate partial models for debugging.
- Extremely extensible: write your own class that subclasses from the in-built
Component
orLikelihood
classes, and it is immediately useable. - Parameters are attached to to the model, for encapsulation, but they can be specified at run-time externally for modularity.
- Models are heirarchical, in the sense that parameters may be specified at any of three levels, and they are propagated through the model heirarchy (note that this doesn't refer to heirarchical parameters, i.e. parameters that depend on other parameters).
- Parameters can be set as fixed or constrained at run-time.
- Models are well-specified, in the sense that they can be entirely specified by a YAML file (and/or written to YAML file), for reproducibility.
This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.
Many of the ideas in this code are adaptations of other MCMC codes, especially CosmoHammer and cobaya.