Overview | Quick install | What does Flax look like? | Documentation
π£ NEW: Check out the NNX API!
This README is a very short intro. To learn everything you need to know about Flax, refer to our full documentation.
Flax was originally started by engineers and researchers within the Brain Team in Google Research (in close collaboration with the JAX team), and is now developed jointly with the open source community.
Flax is being used by a growing community of hundreds of folks in various Alphabet research departments for their daily work, as well as a growing community of open source projects.
The Flax team's mission is to serve the growing JAX neural network research ecosystem -- both within Alphabet and with the broader community, and to explore the use-cases where JAX shines. We use GitHub for almost all of our coordination and planning, as well as where we discuss upcoming design changes. We welcome feedback on any of our discussion, issue and pull request threads. We are in the process of moving some remaining internal design docs and conversation threads to GitHub discussions, issues and pull requests. We hope to increasingly engage with the needs and clarifications of the broader ecosystem. Please let us know how we can help!
Please report any feature requests, issues, questions or concerns in our discussion forum, or just let us know what you're working on!
We expect to improve Flax, but we don't anticipate significant breaking changes to the core API. We use Changelog entries and deprecation warnings when possible.
In case you want to reach us directly, we're at [email protected].
Flax is a high-performance neural network library and ecosystem for JAX that is designed for flexibility: Try new forms of training by forking an example and by modifying the training loop, not by adding features to a framework.
Flax is being developed in close collaboration with the JAX team and comes with everything you need to start your research, including:
-
Neural network API (
flax.linen
): Dense, Conv, {Batch|Layer|Group} Norm, Attention, Pooling, {LSTM|GRU} Cell, Dropout -
Utilities and patterns: replicated training, serialization and checkpointing, metrics, prefetching on device
-
Educational examples that work out of the box: MNIST, LSTM seq2seq, Graph Neural Networks, Sequence Tagging
-
Fast, tuned large-scale end-to-end examples: CIFAR10, ResNet on ImageNet, Transformer LM1b
You will need Python 3.6 or later, and a working JAX installation (with or without GPU support - refer to the instructions). For a CPU-only version of JAX:
pip install --upgrade pip # To support manylinux2010 wheels.
pip install --upgrade jax jaxlib # CPU-only
Then, install Flax from PyPi:
pip install flax
To upgrade to the latest version of Flax, you can use:
pip install --upgrade git https://github.com/google/flax.git
To install some additional dependencies (like matplotlib
) that are required but not included
by some dependencies, you can use:
pip install "flax[all]"
We provide three examples using the Flax API: a simple multi-layer perceptron, a CNN and an auto-encoder.
To learn more about the Module
abstraction, check out our docs, our broad intro to the Module abstraction. For additional concrete demonstrations of best practices, refer to our
guides and
developer notes.
from typing import Sequence
import numpy as np
import jax
import jax.numpy as jnp
import flax.linen as nn
class MLP(nn.Module):
features: Sequence[int]
@nn.compact
def __call__(self, x):
for feat in self.features[:-1]:
x = nn.relu(nn.Dense(feat)(x))
x = nn.Dense(self.features[-1])(x)
return x
model = MLP([12, 8, 4])
batch = jnp.ones((32, 10))
variables = model.init(jax.random.key(0), batch)
output = model.apply(variables, batch)
class CNN(nn.Module):
@nn.compact
def __call__(self, x):
x = nn.Conv(features=32, kernel_size=(3, 3))(x)
x = nn.relu(x)
x = nn.avg_pool(x, window_shape=(2, 2), strides=(2, 2))
x = nn.Conv(features=64, kernel_size=(3, 3))(x)
x = nn.relu(x)
x = nn.avg_pool(x, window_shape=(2, 2), strides=(2, 2))
x = x.reshape((x.shape[0], -1)) # flatten
x = nn.Dense(features=256)(x)
x = nn.relu(x)
x = nn.Dense(features=10)(x)
x = nn.log_softmax(x)
return x
model = CNN()
batch = jnp.ones((32, 64, 64, 10)) # (N, H, W, C) format
variables = model.init(jax.random.key(0), batch)
output = model.apply(variables, batch)
class AutoEncoder(nn.Module):
encoder_widths: Sequence[int]
decoder_widths: Sequence[int]
input_shape: Sequence[int]
def setup(self):
input_dim = np.prod(self.input_shape)
self.encoder = MLP(self.encoder_widths)
self.decoder = MLP(self.decoder_widths (input_dim,))
def __call__(self, x):
return self.decode(self.encode(x))
def encode(self, x):
assert x.shape[1:] == self.input_shape
return self.encoder(jnp.reshape(x, (x.shape[0], -1)))
def decode(self, z):
z = self.decoder(z)
x = nn.sigmoid(z)
x = jnp.reshape(x, (x.shape[0],) self.input_shape)
return x
model = AutoEncoder(encoder_widths=[20, 10, 5],
decoder_widths=[5, 10, 20],
input_shape=(12,))
batch = jnp.ones((16, 12))
variables = model.init(jax.random.key(0), batch)
encoded = model.apply(variables, batch, method=model.encode)
decoded = model.apply(variables, encoded, method=model.decode)
In-detail examples to train and evaluate a variety of Flax models for Natural Language Processing, Computer Vision, and Speech Recognition are actively maintained in the π€ Transformers repository.
As of October 2021, the 19 most-used Transformer architectures are supported in Flax and over 5000 pretrained checkpoints in Flax have been uploaded to the π€ Hub.
To cite this repository:
@software{flax2020github,
author = {Jonathan Heek and Anselm Levskaya and Avital Oliver and Marvin Ritter and Bertrand Rondepierre and Andreas Steiner and Marc van {Z}ee},
title = {{F}lax: A neural network library and ecosystem for {JAX}},
url = {http://github.com/google/flax},
version = {0.9.0},
year = {2024},
}
In the above bibtex entry, names are in alphabetical order, the version number is intended to be that from flax/version.py, and the year corresponds to the project's open-source release.
Flax is an open source project maintained by a dedicated team in Google Research, but is not an official Google product.