torchdyn reference documentation

(New release 1.0 is out!)

torchdyn is a Python library entirely dedicated to continuous, implicit neural architectures and the numerical methods that underpin them.

It features a self-contained numerical suite of differential equation and root solvers, including sensitivity methods (continuous backsolve or interpolated adjoints, reverse-mode AD). The library further contains a model zoo and an extensive set of tutorials for researchers and practitioners.

pip install torchdyn

GitHub link.

Note

• This library is developed and maintained by Michael Poli & Stefano Massaroli, with gracious contributions from the community.

Refer to the links below for a quickstart to core torchdyn features and a description of the library goals and design principles.

A set of extended tutorials, covering everything from models, benchmarks and numerics can be found here.

Getting Started

• Quickstart to torchdyn

General Information

• Contributing to torchdyn and DiffEqML
• FAQ

API documentation

• Public API: torchdyn package