Custom steps

Steps perform work within an SMC iteration, receiving the current particle state and a StepContext. For reporting and run-level control after an iteration, see Custom callbacks.

Derive a custom step from SMCStep and implement apply. It receives the current state, context, and random-number generator, and must return (state, context, output):

from sequince.steps.base import SMCStep


class RecordParticleMean(SMCStep):
    def apply(self, state, context, rng):
        mean = state.particles.mean(axis=0)
        context = context.with_metadata(last_mean=mean)
        return state, context, {"particle_mean": mean}

Use context.shared for read-only configuration and context.with_metadata(...) for current coordination values. Return historical diagnostics as the explicit output so that the runner records them in the corresponding entry of outputs.

For compiled JAX execution, custom steps must also obey the functional interface’s fixed-structure requirements. See JAX and functional execution for the functional step contract and initialization hooks.