Common interfaces for probabilistic programming
Author TuringLang
5 Stars
Updated Last
25 Days Ago
Started In
December 2020


CI IntegrationTest Coverage Status Codecov

A new light-weight package to factor out interfaces and associated APIs for modelling languages for probabilistic programming. High level goals are:

  • Definition of an interface of few abstract types and a small set of functions that will be supported all model and trace types.
  • Provision of some commonly used functionality and data structures, e.g., for managing variable names and traces.

This should facilitate reuse of functions in modelling languages, to allow end users to handle models in a consistent way, and to simplify interaction between different languages and sampler implementations, from very rich, dynamic languages like Turing.jl to highly constrained or simplified models such as GPs, GLMs, or plain log-density problems.

A more short term goal is to start a process of cleanly refactoring and justifying parts of AbstractPPL.jl’s design, and hopefully to get on closer terms with Soss.jl.

See interface draft.