gen-rs is a general-purpose framework for probabilistic programming, written in Rust. gen-rs generalizes many features of other Rust-native probabilistic programming systems such as ferric, and is written at a lower-level than languages such as Gen.jl.
This library is highly-experimental (alpha). It implements the Generative Function Interface [GFI] as specified in the Gen.jl whitepaper and Marco Cusumano-Towner's thesis), and a basic set of inference procedures.
TrieFn DSLGenerate visualizations to visualizations with:
shell
python -m venv venv && activate venv/bin/activate && pip install matplotlib
cargo test && python visualization/visualizer.py
Unlike most modern ML systems, probabilistic programming doesn't require a differentiable likelihood; a fast (possibly parallelized) CPU-bound iterator is often sufficient for inference. This aligns well with Rust's principle of "fearless concurrency". However, most embodied (read: practical) modeling efforts still require extensive parameter tuning and Langevin or Hamiltonian Monte Carlo inference moves, to effectively leverage numerical gradients of the local energy landscape in top-down or supervised data processing.
Despite Rust being a delightful experience to program in, AD support and GPU acceleration is still somewhat shaky (given the lack of first-class Rust-native tensor libraries), limiting these applications.
This project was heavily inspired by GenTL and several more fully-featured projects in the OpenGen ecosystem such as Gen.jl, GenParticleFilters, SMCP3, Bayes3D, and GenJax. It's recommended you check out one of those before deciding whether or not this package meets your requirements.