Penzai: JAX research toolkit for building, editing, and visualizing neural nets

Penzai is an ancient Chinese art of forming miniature trees and landscapes, closely related to the Japanese art of bonsai. In the modern context, Penzai is a JAX library focused on creating legible, functional pytree data structures for neural network models. It offers tools for visualization, modification, and analysis of models, making it ideal for research involving model component reverse-engineering, activation inspection, model surgery, and more. Penzai consists of modular tools such as a neural network library, interactive pretty-printer, swiss-army-knife selector, named axis system, and data effects system. It aims to simplify complex tasks related to neural networks. Additional documentation and tutorials can be found on the official Penzai website.

https://github.com/google-deepmind/penzai

To top