mbi.CliqueVector

class mbi.CliqueVector(domain: Domain, cliques: list[tuple[str, ...]], arrays: dict[tuple[str, ...], Factor])[source]

Bases: object

Manages a collection of factors, each associated with a clique.

This class provides a structure for holding and operating on multiple Factor objects, where each factor corresponds to a clique (a subset of attributes) within a larger domain. It’s particularly useful in the context of graphical models for representing sets of potentials or marginals.

domain

The overall domain that encompasses all cliques.

Type:: Domain

cliques

A list of cliques (tuples of attribute names) for which factors are stored.

Type:: list[Clique]

arrays

A dictionary mapping each clique in cliques to its corresponding Factor object.

Type:: dict[Clique, Factor]

Method generated by attrs for class CliqueVector.

Methods

`__init__`	Method generated by attrs for class CliqueVector.
`abstract`
`apply_sharding`	Apply sharding constraint to each factor in the CliqueVector.
`contract`	Computes a new CliqueVector by projecting this one onto a smaller set of cliques.
`dot`	Computes the dot product between this CliqueVector and another.
`exp`	Applies elementwise exponentiation (jnp.exp) to each factor.
`expand`	Re-expresses this CliqueVector over an expanded set of cliques.
`from_projectable`	Creates a CliqueVector by projecting a data source onto the specified cliques.
`log`	Applies elementwise logarithm (jnp.log) to each factor.
`normalize`	Normalizes each factor within the CliqueVector.
`ones`	Creates a CliqueVector initialized with one factors for each clique.
`parent`	Finds a clique in this vector that is a superset of the given clique.
`project`
`random`	Creates a CliqueVector initialized with random factors for each clique.
`size`	Calculates the total number of parameters across all factors in the vector.
`supports`	Checks if the given clique is supported (is a subset of any clique in the vector).
`zeros`	Creates a CliqueVector initialized with zero factors for each clique.

Attributes

`active_domain`	Returns the merged domain encompassing all attributes across all cliques.
`domain`
`cliques`
`arrays`

domain: Domain

cliques: list[tuple[str, ...]]

arrays: dict[tuple[str, ...], Factor]

classmethod zeros(domain: Domain, cliques: list[tuple[str, ...]]) → CliqueVector[source]: Creates a CliqueVector initialized with zero factors for each clique.

classmethod ones(domain: Domain, cliques: list[tuple[str, ...]]) → CliqueVector[source]: Creates a CliqueVector initialized with one factors for each clique.

classmethod random(domain: Domain, cliques: list[tuple[str, ...]]) → CliqueVector[source]: Creates a CliqueVector initialized with random factors for each clique.

classmethod abstract(domain: Domain, cliques: list[tuple[str, ...]]) → CliqueVector[source]

classmethod from_projectable(data: Projectable, cliques: list[tuple[str, ...]]) → CliqueVector[source]: Creates a CliqueVector by projecting a data source onto the specified cliques.

property active_domain: Domain: Returns the merged domain encompassing all attributes across all cliques.

parent(clique: tuple[str, ...]) → tuple[str, ...] | None[source]: Finds a clique in this vector that is a superset of the given clique.

supports(clique: tuple[str, ...]) → bool[source]: Checks if the given clique is supported (is a subset of any clique in the vector).

project(clique: tuple[str, ...], log: bool = False) → Factor[source]

expand(cliques: list[tuple[str, ...]]) → CliqueVector[source]

Re-expresses this CliqueVector over an expanded set of cliques.

If the original CliqueVector represents the potentials of a Graphical Model, the given cliques support the cliques in the original CliqueVector, then the distribution represented by the new CliqueVector will be identical.

Parameters:: cliques – The new cliques the clique vector will be defined over.
Returns:: An expanded CliqueVector defined over the given set of cliques.

contract(cliques: list[tuple[str, ...]], log: bool = False) → CliqueVector[source]: Computes a new CliqueVector by projecting this one onto a smaller set of cliques.

normalize(total: float = 1, log: bool = True) → CliqueVector[source]: Normalizes each factor within the CliqueVector.

exp() → CliqueVector[source]: Applies elementwise exponentiation (jnp.exp) to each factor.

log() → CliqueVector[source]: Applies elementwise logarithm (jnp.log) to each factor.

dot(other: CliqueVector) → Array | ndarray | bool | number | float | int[source]: Computes the dot product between this CliqueVector and another.

size() → int[source]: Calculates the total number of parameters across all factors in the vector.

apply_sharding(mesh: Mesh | None) → CliqueVector[source]

Apply sharding constraint to each factor in the CliqueVector.

The sharding strategy is automatically determined based on the provided mesh and the factor domains.

Parameters:: mesh – The mesh over which the factors should be sharded.
Returns:: A new CliqueVector identical to self with sharding constraints applied to the underlying factors.