google-deepmind/sonnet

Provides a module-centric architecture centered on `snt.Module` for building self-contained, composable neural network components with lazy parameter initialization based on input shape. Includes built-in modules like Linear, Conv2D, and BatchNorm, plus higher-level networks like MLP, while remaining agnostic to training frameworks—users define their own training loops. Integrates seamlessly with TensorFlow 2's checkpointing, eager execution, and distributed training via `snt.distribute`.

9,907 stars.