ProbZelus

ProbZelus is synchronous probabilistic programming language. It is a conservative extension of Zelus with probabilistic constructs to model uncertainties and perform inference-in-the-loop.

More information about ProbZelus are available in the paper Reactive Probabilistic Programming.

@inproceedings{rppl-plid20,
  author = {Baudart, Guillaume and Mandel, Louis and Atkinson, Eric and Sherman, Benjamin and Pouzet, Marc and Carbin, Michael},
  title = {Reactive Probabilistic Programming},
  year = {2020},
  booktitle = {Proceedings of the 41st ACM SIGPLAN Conference on Programming Language Design and Implementation},
  series = {PLDI 2020}
}

Install

The easiest way to install probzelus is via Opam, the OCaml package manager. You first need to install openblas.

Then pin the packages defined in this repo to add them to opam.

opam pin -k path -n zelus-libs
opam pin -k path -n probzelus

You can now install ProbZelus with

opam install probzelus

This will install a probzeluc executable and the probzelus library in you Opam/OCaml echosystem.

An optional plplot library based on owl-plplot that can be built with:

opam install zelus-owl-plplot zelus-io

Docker

We also provide a docker file to build an image.

make docker_build
make docker_run

A Simple Example

Consider the example of a Hidden Markov Model. The probzelus code is the following

open Probzelus
open Distribution
open Infer_pf

let proba hmm obs = p where
  rec p = sample (gaussian(0.0 fby p, 1.0))
  and () = observe (gaussian(p, 0.5), obs)

let node main () = () where
  rec obs = 0.0 fby (obs +. 1.1)
  and pos_dist = infer 1000 hmm obs
  and mean, std = stats_float pos_dist
  and () =
    print_string " mean: "; print_float mean;
    print_string " std: ";  print_float std;
    print_newline ()

We assume that at each step the position p is not too far from the previous position 0.0 fby p, and that this position is also close to the observation obs.

Node main launches the inference with 1000 particles and print the mean and standard deviation of the inferred distribution at each step. Here the observations are defined with a simple equation: starting from 0.0, at each step we add 1.1 to the value of previous observation.

obs = 0.0, 1.1, 2.2, 3.3, 4.4, 5.5, ...

Compilation

The probzeluc executable is a wrapper around the Zelus compiler. It takes a zelus file (e.g., hmm.zls) and compiles it to OCaml code (e.g., hmm.ml)

probzeluc hmm.zls

You can also specify a simulation node. The compiler then generates an additional files containing the simulation code (e.g., main.ml).

probzeluc -s main hmm.zls

To build an executable, you can then compile the simulation code using the probzelus library.

ocamlfind ocamlc -linkpkg -package probzelus hmm.ml main.ml -o hmm

Other Examples

A set of examples is available in the examples directory. Most of them can be built and executed with:

make exec

The probzelus code is in the *.zls files

Most of the examples requires a X11 server. It can be install on MacOS using XQuartz.

To compile and execute the benchmark, you need the following additional dependencies:

opam install csv mtime

Contributions

ProbZelus is still under development and we welcome contributions. Contributors are expected to submit a 'Developer's Certificate of Origin' which can be found in DCO1.1.txt.