parax 0.9.5

Parametric modeling in JAX

Parax

Parax is a library for parametric modeling in JAX. Features include:

• Parameters with metadata
• PyTrees parameterization via unwrapping
• Derived, constrained, fixed, and random array-like variables
• Abstract interfaces and associated tree manipulation tools

This makes Parax great for:

• Constraints for machine learning
• Bounded optimization for scientific modeling
• Probabilistic modeling and Bayesian inference
• Deep, nested PyTrees
• Combinations of the above

Note that Parax is not a framework, though it can be used to make one. Rather, it is focused on extensibility and interoperability with other JAX libraries (especially Equinox).

Installation

Parax can be installed using pip:

pip install parax

For some built-in constraints and probabilistic features, you may need this distreqx branch:

pip install git+https://github.com/gvcallen/distreqx.git

Documentation

Documentation is available here.

Quick example

Parax provides array-like variables that hold metadata and can be parameterized/constrained:

import parax as prx
import jax.numpy as jnp

p1 = prx.Tagged(1.0, metadata={'hello', 'world'})
p2 = prx.Constrained(prx.constraints.Interval(0.0, 10.0), value=8.0)

p2.raw_value, p2.bounds
# Array(1.3862944), (Array(0.0), Array(10.0))

jnp.sin(p1) + (2 * p2)
# Array(16.84147)

You can also apply arbitrary computations to PyTrees and parameters using explicit unwrapping:

pytree = {'a': 1.0, 'b': {'x': 2.0, 'y': prx.Derived(jnp.log, 3.0)}}
wrapped = prx.Apply(jnp.exp, pytree)

prx.unwrap(wrapped)
# {'a': Array(2.7182817),
#  'b': {'x': Array(7.389056), 
#        'y': Array(3.0)}}

In the above example, prx.Apply operates on the whole PyTree's array-like nodes, while prx.Derived is an array-like prx.AbstractVariable.

Motivation

Usually, PyTrees are just "dumb" containers. However, it is often desirable to attach some metadata/parameterization to a specific node. This can be done by "unwrapping" the metadata or constraint during model preparation or computation.

Compared to other approaches, this provides a middle ground between purity and rigidity:

• The "purist" approach is using shadow PyTrees i.e. parallel trees that hold the relevant metadata/parameterization. However, these are tedious to define for nested models, and require the entire library to manage parallel structures.
• The "standard" approach is using properties and attributes i.e. defining the metadata/parameterization implicitly within the model. This is straight-forward, but tightly couples the extra state with the model, resulting in unnecessary fields and computations.

Next steps

Several more involved examples are available in the documentation, for example on bounded optimization and Bayesian sampling.

Related

The library's design was inspired by several others that deserve mention, including Flax, paramax, and PyTorch.