pyhgf.utils.vectorised_belief_propagation.run_continuous_scan#

pyhgf.utils.vectorised_belief_propagation.run_continuous_scan(network, ys, time_steps, record=())[source]#

Filter a sequence of observations through a continuous network.

Runs jax.lax.scan over (prediction sweep, update sweep) pairs. There is no weight learning: the coupling matrices are parameters of the filter.

Parameters:

  • network (VectorisedNetwork) – The initial continuous network state.

  • ys (Array) – Observations clamped on layer 0 at each step, shape (T, n_obs).

  • time_steps (Array) – Per-step time steps, shape (T,).

  • record (tuple) – Tuple of LayerState field names to record at every step. With an empty tuple (default) the per-step output is layer 0’s expected_mean alone; otherwise it is (traj_step, prediction).

Returns:

  • (final_network, step_output) with per-step outputs stacked along a

  • leading (T,) axis.

Return type:

tuple