Skip to main content

Immutable pytree modules classes with easy manipulation and serialization

Project description

FlareJax

Simple pytree module classes for Jax, strongly inspired by Equinox

  • Referential transparency via strict immutability
  • Safe serialization including hyperparameters
  • Bound methods and function transformations are also modules
  • Auxillary information in key paths for filtered transformations

Quick Examples

Modules work similar to dataclasses, but with the added benefit of being pytrees. Making them compatible with all Jax function transformations.

import flarejax as fj

class Linear(fj.Module):
    # The __init__ method is automatically generated
    w: jax.Array
    b: jax.Array

    # additional intialization methods via classmethods
    @classmethod
    def init(cls, key, dim_in, dim):
        w = jax.random.normal(key, (dim, dim_in)) * 0.02
        b = jax.numpy.zeros((dim,))
        return cls(w=w, b=b)

    def __call__(self, x):
        return self.w @ x + self.b

key = jax.random.PRNGKey(42)
key1, key2 = jax.random.split(key)

model = fj.Sequential(
    (
        Linear.init(key1, 3, 2),
        Linear.init(key2, 2, 5),
    )
)

The model can be serialized and deserialized using fj.save and fj.load.

fj.save("model.npz", model)
model = fj.load("model.npz")

Flarejax includes wrappers of the Jax function transformations, which return callable modules.

model = fj.VMap(model)
model = fj.Jit(model)

Installation

Memmpy can be installed directly from PyPI using pip. It requires Python 3.10+ and Jax 0.4.26+.

pip install flarejax

Design

Flarejax modules sacrifice some flexibility for the sake of a unified interface and safety. Flarejax code should alway be easy to reason about and should not contain any footguns from using python magic.

  1. Everything is immutable and
  2. module fields can be either jax arrays, other modules or json-like data.

This makes it harder to use other jax libraries in flarejax modules. It is recommended to wrap the needed functionality in a module. Most jax libraries should be compatible with flarejax modules, since they are simply callable pytrees.

Roadmap

  • Filtered grad transformation based on key paths
  • Pretty printing for modules
  • Rule to infer static arguments in jitted functions, possibly everything except JAX arrays

See also

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

flarejax-0.3.6.tar.gz (9.0 kB view details)

Uploaded Source

Built Distribution

flarejax-0.3.6-py3-none-any.whl (12.1 kB view details)

Uploaded Python 3

File details

Details for the file flarejax-0.3.6.tar.gz.

File metadata

  • Download URL: flarejax-0.3.6.tar.gz
  • Upload date:
  • Size: 9.0 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/4.0.2 CPython/3.11.9

File hashes

Hashes for flarejax-0.3.6.tar.gz
Algorithm Hash digest
SHA256 7b2fd5302f1814e020fd97d94b907346734eab830bfa63dbb12cd74920cf969f
MD5 f0c7682cb96c59a08045b40df80c9726
BLAKE2b-256 838956b818e9416c77420e6aac37c86a387d2ee156703decb3193377af1b596c

See more details on using hashes here.

File details

Details for the file flarejax-0.3.6-py3-none-any.whl.

File metadata

  • Download URL: flarejax-0.3.6-py3-none-any.whl
  • Upload date:
  • Size: 12.1 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/4.0.2 CPython/3.11.9

File hashes

Hashes for flarejax-0.3.6-py3-none-any.whl
Algorithm Hash digest
SHA256 89973aa1bd6613f13eebc42393bacca9bb886df681e31b71c6332eef036588d9
MD5 c18563596c9d6535e08e7d28d0f64705
BLAKE2b-256 ceead93aab026a5526fe4541ebb04a57ef3131e04e8f21c79bc4378b50759743

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page