jpu 0.0.5rc1

JAX + Units

JAX + Units

Built with JAX and Pint!

This module provides an interface between JAX and Pint to allow JAX to support operations with units. The propagation of units happens at trace time, so jitted functions should see no runtime cost. This library is experimental so expect some sharp edges.

For example:

>>> import jax
>>> import jax.numpy as jnp
>>> import jpu
>>>
>>> u = jpu.UnitRegistry()
>>>
>>> @jax.jit
... def add_two_lengths(a, b):
...     return a + b
...
>>> add_two_lengths(3 * u.m, jnp.array([4.5, 1.2, 3.9]) * u.cm)
<Quantity([3.045 3.012 3.039], 'meter')>

Installation

To install, use pip:

python -m pip install jpu

The only dependencies are jax and pint, and these will also be installed, if not already in your environment. Take a look at the JAX docs for more information about installing JAX on different systems.

Usage

Here is a slightly more complete example:

>>> import jax
>>> import numpy as np
>>> from jpu import UnitRegistry, numpy as jnpu
>>>
>>> u = UnitRegistry()
>>>
>>> @jax.jit
... def projectile_motion(v_init, theta, time, g=u.standard_gravity):
...     """Compute the motion of a projectile with support for units"""
...     x = v_init * time * jnpu.cos(theta)
...     y = v_init * time * jnpu.sin(theta) - 0.5 * g * jnpu.square(time)
...     return x.to(u.m), y.to(u.m)
...
>>> x, y = projectile_motion(
...     5.0 * u.km / u.h, 60 * u.deg, np.linspace(0, 1, 50) * u.s
... )

Technical details & limitations

The most significant limitation of this library is the fact that users must use jpu.numpy functions when interacting with "quantities" with units instead of the jax.numpy interface. This is because JAX does not (yet?) provide a general interface for dispatching of ufuncs on custom array classes. I have played around with the undocumented __jax_array__ interface, but it's not really flexible enough, and it isn't currently compatible with Pytree objects.

So far, only a subset of the numpy/jax.numpy interface is implemented. Pull requests adding broader support (including submodules) would be welcome!