Source code for hybrid_aif.learning.dirichlet_updates
"""Conjugate Dirichlet count updates for discrete observation likelihoods.
The public API is :meth:`~hybrid_aif.distributions.conditional.ConditionalDirichlet.update`
(one modality at a time; the array kernel lives beside it in the distributions
layer). This module keeps the multi-modality convenience wrapper over raw
concentration arrays, one per modality, shaped ``(num_obs_m, *[num_levels_f
for f in A_deps[m]])`` — the same child-first layout as
:mod:`hybrid_aif.inference.discrete`'s likelihood tensors. Collections are
canonically tuples; any sequence is accepted at runtime. The multi-modality wrapper is after
pymdp's ``learning.update_obs_likelihood_dirichlet`` (pymdp/learning.py).
"""
import jax.numpy as jnp
from jax import Array
from hybrid_aif.distributions import Categorical, ConditionalDirichlet
[docs]
def update_likelihood_dirichlet(
prior: tuple[Array, ...],
observations: tuple[Array, ...],
posteriors: tuple[Categorical, ...],
A_deps: tuple[tuple[int, ...], ...],
lr: float = 1.0,
) -> tuple[Array, ...]:
"""One conjugate count update per modality, on positional arrays.
``prior`` holds each modality's child-first concentration table, ``observations`` its
observation (a level index or a probability vector), ``posteriors`` one belief per state
factor and ``A_deps`` each modality's parent state factors; each table gains
``lr * onehot(o) ⊗ q(parents)`` (:meth:`ConditionalDirichlet.update`). Returns the updated
tables in modality order.
"""
updated = []
for conc_m, obs_m, deps_m in zip(prior, observations, A_deps, strict=True):
cd = ConditionalDirichlet(concentration=jnp.asarray(conc_m), parent_ndims=len(deps_m))
beliefs = tuple(posteriors[f] for f in deps_m)
updated.append(cd.update(jnp.asarray(obs_m), beliefs, lr=lr).concentration)
return tuple(updated)