parax 0.5.8

Parametric modeling in JAX

Parax

Parax
Author	Gary Allen
Homepage	github.com/parax/parax
Docs	gvcallen.github.io/parax

Parax is a library for parametric modeling in JAX.

Features

• Derived/constrained parameters with metadata
• Computed PyTrees and callable parameterizations
• Abstract interfaces for fixed, bounded, and probabilistic PyTrees and parameters
• Filtering and manipulation tools

Installation

Parax can be installed using pip:

pip install parax

You may need a custom distreqx branch for some constraints:

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

Overview

Parax aims to provide a foundation for "parametric modeling", i.e. modeling with a focus on the concept of a parameter as a derived array with metadata. This means supporting parameterizations, constraints, bounds, priors, units, and arbitrary metadata, which are needed in both machine learning and scientific modeling.

Parax accomplishes the above in a general manner by providing a common set of abstract interfaces along with filters and tree utilities that use these interfaces. The goal is then to provide a range of tools and concrete classes to minimize boilerplate for users, while still keeping the library extendable and opt-in.

Although Parax can be used in any JAX code, it places emphasis on interoperatibility with Equinox. For example, parax.AbstractConstant and parax.is_constant allow easy partitioning of model parameters using eqx.partition, with parax.Fixed and parax.Frozen providing concrete implementations.

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

Example 1: Constrained Parameters

parax.Param represents a simple JAX array with metadata, while parax.Constrained also caters for built-in constraints. Both classes override parax.AbstractVariable, providing array-like behaviour via __jax_array__. We call variables and/or arrays "param-like".

The example below demonstrates defining a parax.Constrained parameter and then using it in a JAX expression.

import jax.numpy as jnp
import parax as prx

# Define a parameter bounded between 0 and 10
p = prx.Constrained(8.0, prx.Interval(0.0, 10.0))

p.constraint.bounds
# (Array(0., dtype=float32), Array(10., dtype=float32))

# We can use the parameter directly in an equation
jnp.sin(p) + (p * 2.0)
# Array(16.989359, dtype=float32)

# The raw (unconstrained) value used by optimizers under the hood
p.raw_value
# Array(1.3862944, dtype=float32)

# We can also unwrap it explicitly
prx.unwrap(p)
# Array(8., dtype=float32)

Example 2: PyTree Parameterizations

While the above approach caters for array constraints, it is sometimes useful to apply computations over an entire PyTree. To accomplish this, Parax uses unwrapping.

In the following example, we apply jnp.exp to a simple PyTree using parax.Computed and parax.unwrap.

import jax.numpy as jnp
import parax as prx

# Define a PyTree using a dictionary
pytree = {'a': 1.0, 'b': {'x': 10.0, 'y': 20.0}}

# Wrap the PyTree in `parax.Computed`.
wrapped = prx.Computed(pytree, jnp.exp)

# Unwrap the Pytree, applying the computation
prx.unwrap(wrapped)
# {'a': Array(2.7182817, dtype=float32),
#  'b': {'x': Array(22026.465, dtype=float32), 
#        'y': Array(4.851652e+08, dtype=float32)}}

Example 3: Optimizing an eqx.Model using Optimistix

In this example, we define a damped pendulum model using equinox.Module and optimize it using optimistix. The first parameter is initialized with a standard JAX array which we then fix. The second parameter is initialized with an unconstrained prx.Param with dummy metadata. The final parameter is given a default physical scale and constraint during model definition, which we then initialize using a simple float value later.

import jax
import jax.numpy as jnp
import jax.random as jr
import equinox as eqx
import optimistix as optx
import dataclasses
import parax as prx

class DampedPendulum(eqx.Module):
    # Non-default parameters
    k: prx.ParamLike
    friction: prx.ParamLike
    
    # Default physical parameters (creates a prx.Physical)
    length: prx.ParamLike = prx.physical(scale='mm', constraint=prx.Positive())

    def __call__(self, state):
        return self.k * state * self.friction / self.length
    
# Create our model and fix the multiplier
initial_model = DampedPendulum(
    k=jnp.array(1.0),
    friction=prx.Param(0.1, metadata={'hello': 'world'}),
    length=9.81,
)
initial_model = dataclasses.replace(initial_model, k=prx.Fixed(initial_model.k))

# Partition the model, stopping at any constants e.g. `prx.Fixed` variables and `prx.Frozen` layers.
# Then, define the loss function.
params, static = eqx.partition(initial_model, eqx.is_inexact_array, is_leaf=prx.is_constant)

def loss_fn(params, args):
    model = prx.unwrap(eqx.combine(params, static))
    x, y = args
    predictions = jax.vmap(model)(x)
    return jnp.sum((predictions - y)**2)

# Generate some dummy data with friction/length ratio of 0.25
x_data = jnp.linspace(0, 10, 100)
noise = jr.normal(jr.key(42), x_data.shape) * 0.1
y_data = x_data * 0.25 + noise

# Run the optimization
solver = optx.LBFGS(rtol=1e-5, atol=1e-5)
results = optx.minimise(
    fn=loss_fn,
    solver=solver, 
    y0=params, 
    args=(x_data, y_data),
)

# Reconstruct the optimized model
final_model = prx.unwrap(eqx.combine(results.value, static))
final_model.friction # Array(2.4575772, dtype=float32)
final_model.length # Array(9.834487, dtype=float32)

# The optimizer found our ratio of 0.25
final_model.friction / final_model.length 
# Array(0.24989378, dtype=float32)