Haiku

Uses hk.Module subclasses that mirror Sonnet's design, allowing familiar object-oriented neural network construction, as shown in the user manual with examples like MyLinear.

hk.transform converts stateful modules into pure init and apply functions, seamlessly integrating with JAX transforms like jax.jit and jax.grad, demonstrated in the quickstart loss function.

Provides hk.next_rng_key() for deterministic random number generation within transformed functions, easing implementation of stochastic models like dropout or VAEs, as detailed in the working with stochastic models section.

APIs are designed to closely match Sonnet 2, making transition from TensorFlow straightforward, as emphasized in the 'Why Haiku?' section for reducing migration effort.