pyro: a scalable deep probabilistic programming library built on PyTorch

What it solves

Pyro provides a way to build and scale deep probabilistic models. It allows users to represent any computable probability distribution, making it easier to handle uncertainty and complex data patterns in a way that scales to large datasets without the overhead typically found in hand-written code.

How it works

Built on top of PyTorch, Pyro is a deep probabilistic programming language (PPL). It uses a small core of powerful, composable abstractions to allow users to express generative models and inference models. It balances automation with manual control, providing high-level abstractions for those who want automation and direct access for experts who need to customize their inference process.

Who it’s for

It is designed for researchers and developers who need to create flexible, scalable probabilistic models and those who require a balance between high-level automation and low-level control over inference.

Highlights

• Universal: Capable of representing any computable probability distribution.
• Scalable: Designed to handle large datasets with minimal overhead.
• Minimal: Built with a small core of composable abstractions for easier maintenance.
• Flexible: Offers both high-level automation and expert-level customization for inference.