Skip to main content

Inox's banner

Stainless neural networks in JAX

Inox is a minimal JAX library for neural networks with an intuitive PyTorch-like syntax. As with Equinox, modules are represented as PyTrees, which enables complex architectures, easy manipulations, and functional transformations.

Inox aims to be a leaner version of Equinox by only retaining its core features: PyTrees and lifted transformations. In addition, Inox takes inspiration from other projects like NNX and Serket to provide a versatile interface. Despite the differences, Inox remains compatible with the Equinox ecosystem, and its components (modules, transformations, ...) are for the most part interchangeable with those of Equinox.

Inox means "stainless steel" in French 🔪

Installation

The inox package is available on PyPI, which means it is installable via pip.

pip install inox

Alternatively, if you need the latest features, you can install it from the repository.

pip install git+https://github.com/francois-rozet/inox

Getting started

Modules are defined with an intuitive PyTorch-like syntax,

import jax
import inox.nn as nn

init_key, data_key = jax.random.split(jax.random.key(0))

class MLP(nn.Module):
    def __init__(self, key):
        keys = jax.random.split(key, 3)

        self.l1 = nn.Linear(3, 64, key=keys[0])
        self.l2 = nn.Linear(64, 64, key=keys[1])
        self.l3 = nn.Linear(64, 3, key=keys[2])
        self.relu = nn.ReLU()

    def __call__(self, x):
        x = self.l1(x)
        x = self.l2(self.relu(x))
        x = self.l3(self.relu(x))

        return x

model = MLP(init_key)

and are compatible with JAX transformations.

X = jax.random.normal(data_key, (1024, 3))
Y = jax.numpy.sort(X, axis=-1)

@jax.jit
def loss_fn(model, x, y):
    pred = jax.vmap(model)(x)
    return jax.numpy.mean((y - pred) ** 2)

grads = jax.grad(loss_fn)(model, X, Y)

However, if a tree contains strings or boolean flags, it becomes incompatible with JAX transformations. For this reason, Inox provides lifted transformations that consider all non-array leaves as static.

model.name = 'stainless'  # not an array

@inox.jit
def loss_fn(model, x, y):
    pred = jax.vmap(model)(x)
    return jax.numpy.mean((y - pred) ** 2)

grads = inox.grad(loss_fn)(model, X, Y)

Inox also provides a partition mechanism to split the static definition of a module (structure, strings, flags, ...) from its dynamic content (parameters, indices, statistics, ...), which is convenient for updating parameters.

model.mask = jax.numpy.array([1, 0, 1])  # not a parameter

static, params, others = model.partition(nn.Parameter)

@jax.jit
def loss_fn(params, others, x, y):
    model = static(arrays, others)
    pred = jax.vmap(model)(x)
    return jax.numpy.mean((y - pred) ** 2)

grads = jax.grad(loss_fn)(params, others, X, Y)
params = jax.tree_util.tree_map(lambda p, g: p - 0.01 * g, params, grads)

model = static(params, others)

For more information, check out the documentation and tutorials at inox.readthedocs.io.

Contributing

If you have a question, an issue or would like to contribute, please read our contributing guidelines.

Metadata

Release files for inox 0.8.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for inox 0.8.0
File Size Uploaded Provenance Actions
inox-0.8.0.tar.gz 32.0 kB Without provenance Details for inox-0.8.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for inox 0.8.0
File Interpreter ABI Platform Actions
inox-0.8.0-py3-none-any.whl Python 3 none any Details for inox-0.8.0-py3-none-any.whl

Total release size: 68.0 kB

Release files / inox-0.8.0.tar.gz

Download URL inox-0.8.0.tar.gz
Size 32.0 kB
Tags Source
SHA-256 checksum
How to use checksums
0c7fdec657cfb9a7886f97b124ce3c0edb1ff9cbec1ef8cbdb47d9820b868e5e
BLAKE2b-256 checksum
How to use checksums
259b4c98d2010ce2d9749f43e74702157471a34ea9e69de8ebea5bcaadad7c83
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.13.5

Release files / inox-0.8.0-py3-none-any.whl

Download URL inox-0.8.0-py3-none-any.whl
Size 36.1 kB
Tags Python 3
SHA-256 checksum
How to use checksums
2a3a603dd35c5f690693fbc42707508cd36deebb7c6c72d9ba46d02dc73f95c8
BLAKE2b-256 checksum
How to use checksums
59f343d0b69100c4d36f00731ae4ef70ec89a851ff6b21144a0b678f343db9ec
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.13.5

Security for inox 0.8.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page