`glppl`

Translation of lppl probabilistic programs into directed graphical models. Optionally, transiation of graphical models into portable, memory-safe C99 code.

An lppl program to infer the posterior location and scale parameter of a normal distribution could look like this:

double normal_model(record_t<Normal, Gamma>& r, double obs) {
    auto loc = sample(r, "loc", Normal(0.0, 1.0), rng);
    auto scale = sample(r, "scale", Gamma(1.0, 0.5), rng);
    return observe(r, "data", Normal(loc, scale), obs);
}

The glppl version of it looks like this:

graph_node<Normal> normal_model(gr_pair<Normal, Gamma>& gr, double obs) {
    auto loc = sample_g<Normal>(gr, "loc", rng)(0.0, 1.0);
    auto scale = sample_g<Gamma>(gr, "scale", rng)(1.0, 0.5);
    return observe_g<Normal>(gr, "data", obs)(loc, scale);
}

In return for the miniscule amount of extra complexity, you get a graph data structure suitable for static analysis, more specialized inference algorithms, and/or code generation. You can still use any glppl program with lppl inference algorithms and queryers by converting it to a pure sample-based program, e.g. auto f = to_pp(g);, where g is a glppl program and f is an lppl program.

glppl is also useful for doing inference over open-universe probabilistic programs when you intend to deploy inferred models for prediction or forecasting. Using glppl, you can infer the MLE, MAP, or full posterior over graph structures and choose one or more of these structures to deploy. glppl contains alpha functionality to automatically translate your graph probabilistic program into a bespoke memory-safe (zero dynamic allocation) C99 probabilistic graphical model library constructed from ccyan primitives.

As always, read the documentation, get the source, download tagged versions.

⚠️ Versions of glppl greater than v0.1.0 have changed licenses from GPL3 to AGPL3. Please contact us if you are interested in a license exception.

Install

You need to have a recent version of lppl. Running setup.sh from the directory into which you cloned glppl will do the trick.

License etc.

glppl is licensed under the AGPL v3. If you would like a license exception, please contact lppl@davidrushingdewhurst.com.