Conditional#

class hybrid_aif.Conditional[source]#

Bases: Module

Abstract base for conditional families: Markov kernels P(child | parents) with discrete parents.

A conditional holds one child distribution per parent configuration; the parent axes are model structure, not batch axes, and normalization is over the child axis only. The number of parent axes is the static field parent_ndims; parent_shape gives their sizes and batch_shape the axes to their left. condition_on() takes unbatched parent level indices and returns the child distribution at that configuration (batch it with vmap).

Tabular families (ConditionalCategorical, ConditionalDirichlet) store the child axis first, (..., num_levels, *parent_shape) (the pymdp and SPM layout); ConditionalGaussian puts the parent axes before the Gaussian’s own trailing axes. Design: docs/design/conditional-families.md

abstract property batch_shape: tuple[int, ...]#

Leading (batch) axes: leaf shape minus parent and family structure.

abstract property parent_shape: tuple[int, ...]#

Sizes of the parent axes (one per discrete parent).

abstractmethod condition_on(parent_indices)[source]#

The child distribution at one parent configuration (unbatched indices).

__init__()#