Flax is a neural network library built on top of JAX that prioritizes flexibility and ease of use for machine learning research. Developed by Google DeepMind in close collaboration with the JAX team, Flax provides researchers with tools to experiment with novel training approaches by modifying training loops rather than extending framework features. The library is written primarily in Jupyter Notebook and serves both internal Alphabet teams and the broader JAX research community.
The project released Flax NNX in 2024 as a simplified API designed to make neural network creation, inspection, debugging, and analysis more straightforward. Flax NNX introduces first-class support for Python reference semantics, allowing users to express models using regular Python objects with reference sharing and mutability. This represents an evolution from the Flax Linen API, which was released in 2020 and remains available with its own documentation site. The team has committed to avoiding significant breaking changes to the core API, using changelog entries and deprecation warnings when modifications are necessary.
The neural network API includes standard components such as Linear and Conv layers, normalization techniques like BatchNorm, LayerNorm, and GroupNorm, attention mechanisms including MultiHeadAttention, recurrent cells such as LSTMCell and GRUCell, and regularization through Dropout. Beyond core layers, Flax provides utilities and patterns for replicated training, serialization and checkpointing, metrics tracking, and device-level prefetching. Educational examples demonstrate practical usage through MNIST tutorials and inference with the Gemma language model transformer.
The repository overlaps contributors with github/gh-aw, solo-io/gloo, and longhorn/longhorn projects.
Flax is classified across multiple domains including functional programming, modeling frameworks, automatic differentiation, customization capabilities, transformer models, model training, deep learning, and scalable machine learning. The library emphasizes PyTorch-like API design while leveraging JAX's functional programming paradigm, supporting research use cases including federated learning and tensor operations. Installation requires Python 3.8 or later and JAX, with optional dependencies available for additional functionality like matplotlib. The team actively coordinates through GitHub discussions, issues, and pull requests, welcoming feature requests and community feedback at [email protected].
