This release is a pre-release and may not be stable for production use.
REAX
Scalable, flexible training for JAX, inspired by the simplicity of PyTorch Lightning.
REAX is a lightweight framework for training JAX models. It takes care of the training loop, logging, checkpointing and scaling across devices, so that your code can focus on the model and the science. 🚀
Why REAX? 🌟
Bring your own model 🔧: REAX works with any JAX library, e.g. Flax Linen, Flax NNX or Equinox. Your model stays plain JAX code.
Minimal boilerplate 🧩: write a reax.Module with your training step and optimiser, and let the reax.Trainer run fitting, validation, testing and prediction.
Batteries included 💾: checkpointing, early stopping, progress bars, metrics, and loggers for CSV, TensorBoard and MLflow.
Scales with JAX ⚡: run on multiple devices and processes with minimal code changes.
Stay in control 🎯: hook into every stage of training with module hooks and listeners, or drop down to the reax.Engine to write your own training loop.
Installation 🛠️
With pip:
pip install reax
or with uv, to add REAX to your project:
uv add reax
Example
Train an autoencoder on MNIST, using a Flax Linen model:
from functools import partial
from flax import linen
import jax
import optax
import reax
from reax.demos import mnist
class Autoencoder(linen.Module):
def setup(self):
self.encoder = linen.Sequential([linen.Dense(128), linen.relu, linen.Dense(3)])
self.decoder = linen.Sequential([linen.Dense(128), linen.relu, linen.Dense(28 * 28)])
def __call__(self, x):
return self.decoder(self.encoder(x))
# --------------------------------
# Step 1: Define a REAX Module
# --------------------------------
# A reax.Module defines a full *system* (e.g. an LLM, diffusion model, autoencoder or
# image classifier): the model, how to train it and how to optimise it.
class ReaxAutoEncoder(reax.Module):
def __init__(self):
super().__init__()
self.ae = Autoencoder()
def configure_model(self, stage: reax.Stage, batch, /):
if self.parameters() is None:
# Initialise the parameters using an example batch
x, _ = batch
params = self.ae.init(self.rngs(), x.reshape(x.shape[0], -1))
self.set_parameters(params)
def __call__(self, x):
return self.ae.apply(self.parameters(), x)
def training_step(self, batch, batch_idx):
x, _ = batch
x = x.reshape(x.shape[0], -1)
loss, grads = jax.value_and_grad(self.loss_fn)(self.parameters(), x, self.ae.apply)
self.log("train_loss", loss, on_step=True, prog_bar=True)
# Return the gradients and REAX takes care of the optimiser update
return loss, grads
def validation_step(self, batch, batch_idx):
x, _ = batch
x = x.reshape(x.shape[0], -1)
self.log("val_loss", self.loss_fn(self.parameters(), x, self.ae.apply), prog_bar=True)
@staticmethod
@partial(jax.jit, static_argnums=2)
def loss_fn(params, x, apply_fn):
predictions = apply_fn(params, x)
return optax.losses.squared_error(predictions, x).mean()
def configure_optimizers(self):
opt = optax.adam(learning_rate=1e-3)
return opt, opt.init(self.parameters())
# -------------------
# Step 2: Define data
# -------------------
trainer = reax.Trainer()
dataset = mnist.MnistDataset(download=True)
train, val = reax.data.random_split(trainer.rngs, dataset, [55_000, 5_000])
# -------------------
# Step 3: Train
# -------------------
autoencoder = ReaxAutoEncoder()
trainer.fit(
autoencoder,
reax.ReaxDataLoader(train, batch_size=64, shuffle=True),
reax.ReaxDataLoader(val, batch_size=64),
max_epochs=5,
)
More examples, including Flax NNX and Equinox models, are in the documentation.
Coming from PyTorch Lightning?
REAX ports the abstractions and best practices of PyTorch Lightning to JAX, and parts of its code are adapted directly from Lightning. If you know Lightning, the Trainer, Module, hooks, listeners (Lightning’s callbacks) and loggers will feel familiar. Where it matters, REAX departs from Lightning to fit JAX: parameters and optimiser state are explicit, and training steps return gradients rather than relying on loss.backward().
Metadata
Release files for reax 0.7.0.dev2
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| reax-0.7.0.dev2.tar.gz | 151.1 kB | Without provenance | Details for reax-0.7.0.dev2.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| reax-0.7.0.dev2-py3-none-any.whl | Python 3 | none | any | Details for reax-0.7.0.dev2-py3-none-any.whl |
Total release size: 344.8 kB
Release files / reax-0.7.0.dev2.tar.gz
| Download URL | reax-0.7.0.dev2.tar.gz |
|---|---|
| Size | 151.1 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
6c8c056749ed6b2943e6ba8bd58891c05c5e61eabd6802b940d1d7b178271af8
|
|
BLAKE2b-256 checksum How to use checksums |
b8ed1d6d1e50b555f43c3199054772ad8ee96bd047d4dbcf2346ba9bb3618fc0
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.15
|
Release files / reax-0.7.0.dev2-py3-none-any.whl
| Download URL | reax-0.7.0.dev2-py3-none-any.whl |
|---|---|
| Size | 193.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
14b4911eae1aed357f04f7794f1ed6e6ca9557c4d34215e3c1aab955e6123b0b
|
|
BLAKE2b-256 checksum How to use checksums |
3b7ea315f2ec19fb4ef5fb8219674161469941cd89b7f80c2c628605b4f2ca8d
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.15
|
Security for reax 0.7.0.dev2
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.