TF2JAX
An experimental library that converts TensorFlow functions and graphs into JAX functions for reuse and fine-tuning within JAX codebases.
What is TF2JAX?
TF2JAX is an experimental library that converts TensorFlow functions and computational graphs into equivalent JAX functions. It solves the problem of reusing existing TensorFlow models and code within JAX ecosystems, enabling developers to apply JAX's powerful transformations like automatic differentiation and vectorization to TensorFlow-based workflows.
Machine learning researchers and engineers who work with both TensorFlow and JAX, particularly those looking to migrate TensorFlow models to JAX or integrate TensorFlow components into JAX pipelines.
TF2JAX provides a transparent conversion process that maintains debuggability while enabling full interoperability between TensorFlow and JAX, unlike alternative approaches that treat converted functions as black boxes. It supports important serialization formats and allows fine-tuning of converted models within JAX codebases.
Overview
TF2JAX is an experimental library that transforms TensorFlow functions and graphs into equivalent JAX functions. This enables developers to reuse existing TensorFlow models, including SavedModel and TensorFlow Hub modules, within JAX codebases while maintaining the ability to apply JAX transforms.
Key Features
- TensorFlow to JAX Conversion — Converts
tf.functionand TensorFlow graphs into pure JAX functions that can be used with JAX transforms likejit,grad, andvmap. - Serialization Format Support — Works with TensorFlow SavedModel and legacy TF-Hub formats, allowing conversion directly from serialized models.
- JAX to TensorFlow Roundtrip — Can convert functions serialized via
jax2tf.convertback into JAX, enabling full interoperability between the two frameworks. - Randomness Handling — Properly converts TensorFlow random ops to JAX functions that accept PRNG keys as keyword arguments.
- Custom Gradient Support — Experimental support for converting TensorFlow custom gradients to JAX.
- Configurable Behavior — Offers global and local configuration options for shape/dtype checking, constant precision, and gradient prevention.
Philosophy
TF2JAX aims to provide a transparent conversion pathway that maintains debuggability and introspection while enabling seamless integration between TensorFlow and JAX ecosystems.
Related Projects
Found a gem we're missing?
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
