sonnet

The library's central concept is the snt.Module class a composable building block that manages its own trainable variables and can be nested, reused, and serialized in a clean, object-oriented pattern.

Sonnet was developed internally at DeepMind to support research on complex model architectures including graph neural networks, neural processes, and world models.

The library is particularly valued for its clean implementation of complex architectural patterns attention mechanisms, relational memory, graph networks, and generative model components that researchers reference when implementing or verifying DeepMind's published architectures.

Each module is designed to be stateless in its computation graph construction, making models built with Sonnet easier to trace, export, and analyze than equivalent implementations using TensorFlow's lower-level APIs.

Sonnet is open-source under the Apache 2.0 license and available via pip for TensorFlow 2.x.

While it has been superseded in parts of the research community by JAX-based frameworks like Haiku and Flax (also from DeepMind), Sonnet remains widely used for TensorFlow-based research reproduction and as a reference implementation for DeepMind's published model architectures.

It is particularly useful for researchers implementing papers from DeepMind's publication library who want official, well-tested implementations of the architectural components.