sonnet

Why we included this project

Sonnet was built by DeepMind researchers who use it for their own experiments, so it carries the weight of real research work rather than a polished demo. Everything in the library hangs off one concept, snt.Module, with ready-made pieces like Linear, Conv2D, BatchNorm, and a prebuilt MLP. But it ships no training loop: Sonnet stays deliberately unopinionated, and the modules are self-contained and decoupled, so you can snap custom architectures together without fighting the framework. For teams doing reinforcement learning, distributed training, or anything where you want to see and control exactly what each layer does, that readability and control usually beats a batteries-included framework.