Thinc

Thinc is a lightweight, type checked deep learning library for composing models in Python. It comes from the makers of spaCy and Prodigy, and earlier versions of the library have powered spaCy since its release. The current design is a redesign that takes a functional approach to defining neural networks. The library is framework agnostic. Layers defined in PyTorch, TensorFlow and MXNet can be wrapped and combined in one model, and zero copy array interchange lets developers mix frameworks without changing the application code around them. Thinc ships with type checking so that, for example, passing a one dimensional array into a model that expects two dimensions is flagged as a type error in the editor before the code runs. A configuration system lets developers describe trees of objects with references to their own functions, which replaces passing around loose blobs of settings and is meant to work for both research and production. The library is small, has very few required dependencies, and installs through pip and conda on Linux, macOS and Windows. Thinc is aimed at machine learning engineers and researchers who want a clean model composition layer on top of the major frameworks. The site publishes no pricing, and the library is distributed through the standard Python package channels.