Probabilistic programming with programmable inference for parallel accelerators.

This is the community edition of GenJAX, a probabilistic programming language in development at MIT's Probabilistic Computing Project. We recommend this version for stability, community contributions, expanded features and more active community-driven maintenance. The research version is more likely to be unstable, and evolve sporadically.

🔎 What is GenJAX?

Gen is a multi-paradigm (generative, differentiable, incremental) language for probabilistic programming focused on generative functions: computational objects which represent probability measures over structured sample spaces.

GenJAX is an implementation of Gen on top of JAX - exposing the ability to programmatically construct and manipulate generative functions, as well as JIT compile + auto-batch inference computations using generative functions onto GPU devices.

[!TIP] GenJAX is part of a larger ecosystem of probabilistic programming tools based upon Gen. Explore more...

Quickstart

To install GenJAX, run

bash pip install genjax

Then install JAX using this guide to choose the command for the architecture you're targeting. To run GenJAX without GPU support:

sh pip install jax[cpu]~=0.4.24

On a Linux machine with a GPU, run the following command:

sh pip install jax[cuda12]~=0.4.24

Quick example

The following code snippet defines a generative function called beta_bernoulli that

• takes a shape parameter beta
• uses this to create and draw a value p from a Beta distribution
• Flips a coin that returns 1 with probability p, 0 with probability 1-p and returns that value

Then, we create an inference problem (by specifying a posterior target), and utilize sampling importance resampling to give produce single sample estimator of p.

We can JIT compile that entire process, run it in parallel, etc - which we utilize to produce an estimate for p over 50 independent trials of SIR (with K = 50 particles).

```python import jax import jax.numpy as jnp import genjax from genjax import beta, flip, gen, Target, ChoiceMap from genjax.inference.smc import ImportanceK

Create a generative model.

@gen def beta_bernoulli(α, β): p = beta(α, β) @ "p" v = flip(p) @ "v" return v

@jax.jit def runinference(obs: bool): # Create an inference query - a posterior target - by specifying # the model, arguments to the model, and constraints. posteriortarget = Target(beta_bernoulli, # the model (2.0, 2.0), # arguments to the model ChoiceMap.d({"v": obs}), # constraints )

# Use a library algorithm, or design your own - more on that in the docs!
alg = ImportanceK(posterior_target, k_particles=50)

# Everything is JAX compatible by default.
# JIT, vmap, to your heart's content.
key = jax.random.key(314159)
sub_keys = jax.random.split(key, 50)
_, p_chm = jax.vmap(alg.random_weighted, in_axes=(0, None))(
    sub_keys, posterior_target
)

# An estimate of `p` over 50 independent trials of SIR (with K = 50 particles).
return jnp.mean(p_chm["p"])

(runinference(True), runinference(False)) ```

python (Array(0.6039314, dtype=float32), Array(0.3679334, dtype=float32))

References

Many bits of knowledge have gone into this project -- you can find many of these bits at the MIT Probabilistic Computing Project page under publications. Here's an abbreviated list of high value references:

• Marco Cusumano-Towner's thesis on Gen
• The main Gen.jl repository
• (Trace types) (Lew et al) trace types
• (RAVI) (Lew et al) Recursive auxiliary-variable inference
• (GenSP) Alex Lew's Gen.jl implementation of GenSP
• (ADEV) (Lew & Huot, et al) Automatic differentiation of expected values of probabilistic programs

JAX influences

This project has several JAX-based influences. Here's an abbreviated list:

• This notebook on static dispatch (Dan Piponi)
• Equinox (Patrick Kidger's work on neural networks via callable Pytrees)
• Oryx (interpreters and interpreter design)

Acknowledgements

The maintainers of this library would like to acknowledge the JAX and Oryx maintainers for useful discussions and reference code for interpreter-based transformation patterns.

Disclaimer

This is a research project. Expect bugs and sharp edges. Please help by trying out GenJAX, reporting bugs, and letting us know what you think!

Get Involved + Get Support

Pull requests and bug reports are always welcome! Check out our Contributor's Guide for information on how to get started contributing to GenJAX.

The TL;DR; is:

• send us a pull request,
• iterate on the feedback + discussion, and
• get a +1 from a maintainer

in order to get your PR accepted.

Issues should be reported on the GitHub issue tracker.

If you want to discuss an idea for a new feature or ask us a question, discussion occurs primarily in the body of Github Issues