Inox

Inox is a minimal JAX library for neural networks with an intuitive PyTorch-like interface. As with Equinox, modules are represented as PyTrees, which allows to pass networks in and out of native JAX transformations, like jax.jit or jax.vmap. However, Inox implements modules in a different way such that non-array leaves, like hyper-parameters or boolean flags, are automatically marked as static. Consequently, working with Inox modules does not require custom versions of JAX transformations that filter out non-array leaves.

As a philosophical successor to Equinox, Inox inherits its speed, elegance and expressiveness, but tries to be closer to PyTorch in terms of ease of use and intuitiveness. In short, the strengths of JAX and PyTorch, without the rust.

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

Networks are defined using an intuitive PyTorch-like syntax,

import jax
import inox.nn as nn

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

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

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

network = MLP(init_key)

and are fully compatible with native JAX transformations.

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

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

loss = loss_fn(network, X, Y)