How to create a custom distribution in Distrax?

Subclass distrax.Distribution and implement key methods like _sample_n and log_prob; the README provides a clear example with minimal code, making it accessible for custom probabilistic models.

Distrax vs TensorFlow Probability for JAX projects?

Distrax excels in JAX-native performance and readability but lacks TFP's advanced features. Choose Distrax for lightweight, extensible JAX work; TFP for comprehensive statistical libraries with JAX substrates.

Can I use Distrax with vmap for batched sampling?

Use caution—vmap support is experimental. Simple distributions may work, but complex ones like MultivariateNormalDiag might fail, as noted in the README warnings about incomplete serialization logic.

How to migrate from TensorFlow Probability to Distrax?

Leverage Distrax's cross-compatibility; pass TFP distributions directly to Distrax functions or use to_tfp for wrapping, allowing gradual integration without rewriting entire codebases.

What's the advantage of sample_and_log_prob in Distrax?

It combines sampling and log probability computation in one call, which can be more efficient for certain distributions, reducing overhead and improving performance in JAX transformations.

Is Distrax production-ready for machine learning models?

While efficient and extensible, Distrax's experimental vmap support and subset nature may make it less stable for production compared to mature libraries like TFP, especially for complex use cases.

Distrax — JAX Probability Distributions Library

What is Distrax?

Distrax is a lightweight probability library built for JAX that provides implementations of probability distributions and bijectors. It reimplements a subset of TensorFlow Probability with a focus on JAX-native performance, readability, and extensibility, solving the need for efficient probabilistic computation within the JAX ecosystem.

Target Audience

Researchers and developers working with probabilistic models in JAX, especially those needing custom distributions, bijectors, or compatibility with TensorFlow Probability in JAX-based projects.

Value Proposition

Developers choose Distrax for its JAX-native design, which ensures seamless integration with JAX transformations, its emphasis on code readability and extensibility for custom probabilistic components, and its cross-compatibility with TensorFlow Probability for flexible workflows.

Overview

Distrax is a lightweight probability library built for JAX, offering distributions and bijectors as a reimplementation of core TensorFlow Probability functionality. It provides a JAX-native alternative that emphasizes readability, extensibility, and cross-compatibility with TFP, making it ideal for probabilistic modeling in JAX-based research and applications.

Key Features

JAX-native implementation — Designed from the ground up for JAX, enabling seamless integration with JAX transformations like jit, vmap, and grad.
Cross-compatibility with TensorFlow Probability — Distrax distributions and bijectors can interoperate with TFP equivalents, allowing mixed usage and easy migration.
Extensible architecture — Simple subclassing interfaces for defining custom distributions and bijectors, useful for reinforcement learning policies or novel probabilistic models.
Efficient combined operations — Methods like sample_and_log_prob and forward_and_log_det provide optimized one-line sampling with log-probabilities.
Readable code — Implementations are self-contained and closely follow the underlying mathematics for clarity and ease of understanding.

Philosophy

Distrax prioritizes readability and extensibility while maintaining compatibility with TensorFlow Probability, aiming to be a practical, JAX-friendly tool for probabilistic computation without replicating TFP's full feature set.

Distrax

What is Distrax?

Overview

Key Features

Philosophy

Related Projects

Found a gem we're missing?

Use Cases

Best For

Not Ideal For

Pros & Cons

Pros

Cons

Frequently Asked Questions