compiled-knowledge 4.0.0a20__cp313-cp313-win_amd64.whl

ck/pgm_compiler/recursive_conditioning.py

@@ -0,0 +1,231 @@
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Iterable, Dict, Optional, List, Sequence, Tuple, Set
+
+from ck.circuit import Circuit, CircuitNode
+from ck.pgm import PGM
+from ck.pgm_circuit import PGMCircuit
+from ck.pgm_compiler.support import clusters
+from ck.pgm_compiler.support.circuit_table import CircuitTable
+from ck.pgm_compiler.support.clusters import ClusterAlgorithm
+from ck.pgm_compiler.support.factor_tables import make_factor_tables, FactorTables
+from ck.utils.iter_extras import combos
+
+# Standard cluster algorithms.
+MIN_DEGREE: ClusterAlgorithm = clusters.min_degree
+MIN_FILL: ClusterAlgorithm = clusters.min_fill
+MIN_DEGREE_THEN_FILL: ClusterAlgorithm = clusters.min_degree_then_fill
+MIN_FILL_THEN_DEGREE: ClusterAlgorithm = clusters.min_fill_then_degree
+MIN_WEIGHTED_DEGREE: ClusterAlgorithm = clusters.min_weighted_degree
+MIN_WEIGHTED_FILL: ClusterAlgorithm = clusters.min_weighted_fill
+MIN_TRADITIONAL_WEIGHTED_FILL: ClusterAlgorithm = clusters.min_traditional_weighted_fill
+
+
+def compile_pgm(
+        pgm: PGM,
+        const_parameters: bool = True,
+        *,
+        algorithm: ClusterAlgorithm = MIN_FILL_THEN_DEGREE,
+        pre_prune_factor_tables: bool = False,
+) -> PGMCircuit:
+    """
+    Compile the PGM to an arithmetic circuit, using recursive conditioning.
+
+    Conforms to the `PGMCompiler` protocol.
+
+    Args:
+        pgm: The PGM to compile.
+        const_parameters: If true, the potential function parameters will be circuit
+            constants, otherwise they will be circuit variables.
+        algorithm: algorithm to get an elimination order.
+        pre_prune_factor_tables: if true, then heuristics will be used to remove any provably zero row.
+
+    Returns:
+        a PGMCircuit object.
+    """
+    elimination_order: Sequence[int] = algorithm(pgm).eliminated
+    factor_tables: FactorTables = make_factor_tables(
+        pgm=pgm,
+        const_parameters=const_parameters,
+        multiply_indicators=True,
+        pre_prune_factor_tables=pre_prune_factor_tables,
+    )
+
+    if pgm.number_of_factors == 0:
+        # Deal with special case: no factors
+        top: CircuitNode = factor_tables.circuit.const(1)
+    else:
+        dtree: _DTree = _make_dtree(elimination_order, factor_tables)
+        states: List[Sequence[int]] = [tuple(range(len(rv))) for rv in pgm.rvs]
+        top: CircuitNode = dtree.make_circuit(states, factor_tables.circuit)
+
+    top.circuit.remove_unreachable_op_nodes(top)
+
+    return PGMCircuit(
+        rvs=tuple(pgm.rvs),
+        conditions=(),
+        circuit_top=top,
+        number_of_indicators=factor_tables.number_of_indicators,
+        number_of_parameters=factor_tables.number_of_parameters,
+        slot_map=factor_tables.slot_map,
+        parameter_values=factor_tables.parameter_values,
+    )
+
+
+def _make_dtree(elimination_order: Sequence[int], factor_tables: FactorTables) -> _DTree:
+    if len(factor_tables.tables) == 0:
+        return _DTreeLeaf(CircuitTable(factor_tables.circuit, (), ()))
+
+    # Populate `trees` with all the leaves
+    trees: List[_DTree] = [_DTreeLeaf(table) for table in factor_tables.tables]
+
+    # join trees by elimination random variable
+    for rv_index in elimination_order:
+        next_trees: List[_DTree] = []
+        to_join: List[_DTree] = []
+        for tree in trees:
+            if rv_index in tree.vars:
+                to_join.append(tree)
+            else:
+                next_trees.append(tree)
+        if len(to_join) >= 2:
+            while len(to_join) > 1:
+                # join the two shallowest trees
+                to_join.sort(key=lambda t: -t.depth())
+                x = to_join.pop()
+                y = to_join.pop()
+                to_join.append(_DTreeInterior(x, y))
+            next_trees.append(to_join[0])
+            trees = next_trees
+
+    # Make sure there is only one tree
+    while len(trees) > 1:
+        x = trees.pop(0)
+        y = trees.pop(0)
+        trees.append(_DTreeInterior(x, y))
+
+    root = trees[0]
+    root.update_cutset()
+    return root
+
+
+class _DTree(ABC):
+    """
+    A node in a binary decomposition tree.
+
+    A node is either a _DTreeLeaf, holding a single CircuitTable,
+    or is a _DTreeInterior, which has exactly two children.
+    """
+
+    def __init__(self, vars_idxs: Set[int]):
+        self.vars: Set[int] = vars_idxs
+        self.cutset: Sequence[int] = ()
+        self.context: Sequence[int] = ()
+
+    @abstractmethod
+    def update_cutset(self, acutset: Iterable[int] = ()) -> None:
+        """
+        After the d-tree is defined, call `update_cutset` on the root
+        to ensure all fields are properly set.
+        """
+        ...
+
+    @abstractmethod
+    def make_circuit(self, states: List[Sequence[int]], circuit: Circuit) -> CircuitNode:
+        """
+        After the d-tree is defined and cutsets are updated,
+        construct a circuit using recursive conditioning.
+        """
+        ...
+
+    @abstractmethod
+    def depth(self) -> int:
+        """
+        Tree depth.
+        """
+
+
+@dataclass
+class _DTreeLeaf(_DTree):
+
+    def __init__(self, table: CircuitTable):
+        super().__init__(set(table.rv_idxs))
+        self.table: CircuitTable = table
+
+    def update_cutset(self, acutset: Iterable[int] = ()) -> None:
+        pass
+
+    def make_circuit(self, states: List[Sequence[int]], circuit: Circuit) -> CircuitNode:
+        table = self.table
+
+        key_states: List[Sequence[int]] = [
+            states[rv_idx]
+            for rv_idx in table.rv_idxs
+        ]
+        to_sum: List[CircuitNode] = list(
+            filter(
+                (lambda n: n is not None),
+                (table.get(key) for key in combos(key_states))
+            )
+        )
+        return circuit.optimised_add(to_sum)
+
+    def depth(self) -> int:
+        return 1
+
+
+@dataclass
+class _DTreeInterior(_DTree):
+
+    def __init__(self, x: _DTree, y: _DTree):
+        super().__init__(x.vars.union(y.vars))
+        self.x: _DTree = x
+        self.y: _DTree = y
+        self.cache: Dict[Tuple[int, ...], CircuitNode] = {}
+
+    def update_cutset(self, acutset: Iterable[int] = ()) -> None:
+        cutset: Set[int] = self.x.vars.intersection(self.y.vars).difference(acutset)
+        self.cutset = tuple(cutset)
+        self.context = tuple(self.vars.intersection(acutset))
+
+        next_acutset = cutset.union(acutset)
+        self.x.update_cutset(next_acutset)
+        self.y.update_cutset(next_acutset)
+
+    def make_circuit(self, states: List[Sequence[int]], circuit: Circuit) -> CircuitNode:
+
+        assert all(len(states[rv_idx]) == 1 for rv_idx in self.context), 'consistency check'
+        context_key: Tuple[int, ...] = tuple(
+            states[rv_idx][0]
+            for rv_idx in self.context
+        )
+
+        cache: Optional[CircuitNode] = self.cache.get(context_key)
+        if cache is not None:
+            return cache
+
+        cutset = self.cutset
+        key_states: List[Sequence[int]] = [
+            states[rv_idx]
+            for rv_idx in cutset
+        ]
+        to_sum: List[CircuitNode] = []
+        for key in combos(key_states):
+            # Update the evidence with the keys
+            next_states = states.copy()
+            for s, i in zip(key, cutset):
+                next_states[i] = (s,)
+
+            # Recursively call
+            x_node = self.x.make_circuit(next_states, circuit)
+            y_node = self.y.make_circuit(next_states, circuit)
+            to_sum.append(circuit.optimised_mul((x_node, y_node)))
+
+        result = circuit.optimised_add(to_sum)
+        self.cache[context_key] = result
+        return result
+
+    def depth(self) -> int:
+        return max(self.x.depth(), self.y.depth())

ck/pgm_compiler/support/circuit_table/__init__.py

@@ -0,0 +1,17 @@
+# There are two implementations of the `circuit_table` module are provided
+# for developer R&D purposes. One is pure Python and the other is Cython.
+# Which implementation is used can be selected here.
+#
+# A similar selection can be made for the `circuit` module.
+# Note that if the Cython implementation is chosen for `circuit_table` then
+# the Cython implementation must be chosen for `circuit`.
+
+# from ._circuit_table_py import (
+from ._circuit_table_cy import (
+    CircuitTable,
+    TableInstance,
+    sum_out,
+    sum_out_all,
+    project,
+    product,
+)