compiled-knowledge 4.1.0a2__cp312-cp312-macosx_11_0_arm64.whl → 4.1.0a3__cp312-cp312-macosx_11_0_arm64.whl

ck/learning/coalesce_cross_tables.py

@@ -0,0 +1,395 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Tuple, List, Sequence, Dict
+
+import numpy as np
+from scipy.sparse import dok_matrix
+from scipy.sparse.linalg import lsqr
+
+from ck.dataset.cross_table import CrossTable
+from ck.pgm import RandomVariable, Instance
+from ck.utils.iter_extras import combos
+from ck.utils.np_extras import NDArrayFloat64
+
+
+def coalesce_cross_tables(crosstabs: Sequence[CrossTable], rvs: Sequence[RandomVariable]) -> CrossTable:
+    """
+    Rationalise multiple cross-tables into a single cross-table.
+
+    This method implements a solution finding the best vector `a`
+    that solves `b = m a`, subject to `a[i] >= 0`.
+
+    `a` is a column vector with one entry for each instance of rvs seen in the cross-tables.
+
+    `b` is a column vector containing all probabilities from all cross-tables (normalised weights).
+
+    `m` is a sparse matrix with m[i, j] = 1 where b[j] is in the sum for source probability a[i].
+
+    Args:
+        crosstabs: a collection of cross-tables to coalesce.
+        rvs: the random variables that cross-tables will be projected on to.
+
+    Returns:
+        a cross-table defined for the given `rvs`, with values inferred from `crosstabs`
+    """
+    if len(crosstabs) == 0:
+        return CrossTable(rvs)
+
+    m: dok_matrix
+    b: np.ndarray
+    a_keys: Sequence[Instance]
+    m, b, a_keys = _make_matrix(crosstabs, rvs)
+
+    a = _solve(m, b)
+
+    return CrossTable(
+        rvs,
+        update=(
+            (instance, weight)
+            for instance, weight in zip(a_keys, a)
+        ),
+    )
+
+
+def _solve(m: dok_matrix, b: np.ndarray) -> np.ndarray:
+    """
+    Find the best 'a' for `b = m a` subject to `a[i] >= 0`.
+    """
+    assert len(b.shape) == 1, 'b should be a vector'
+    assert b.shape[0] == m.shape[0], 'b and m must be compatible'
+
+    return _solve_sam(m, b)
+    # return _solve_lsqr(m, b)
+    # return _solve_pulp_l1(m, b)
+
+
+def _solve_sam(m: dok_matrix, b: np.ndarray) -> np.ndarray:
+    """
+    Find the best 'a' for `b = m a` subject to `a[i] >= 0`.
+
+    Uses a custom 'split and mean' (SAM) method.
+    """
+    sam = _SAM(m, b)
+    a, error = sam.solve(
+        max_iterations=100,
+        tolerance=1e-6,
+        change_tolerance=1e-12
+    )
+    return a
+
+
+def _solve_lsqr(m: dok_matrix, b: np.ndarray) -> np.ndarray:
+    """
+    Find the best 'a' for `b = m a` subject to `a[i] >= 0`.
+
+    Uses scipy `lsqr` method, with a heuristic to fix negative values in `a`.
+    """
+    a: np.ndarray
+    # Pycharm type checker incorrectly infers the type signature of `lsqr`
+    # noinspection PyTypeChecker
+    a, istop, itn, r1norm, r2norm, _, _, _, _, _ = lsqr(m, b)
+
+    # Negative values or values > 1 are not a valid solution.
+
+    # Heuristic fix up...
+    if len(a) > 0:
+        min_val = np.min(a)
+        if min_val < 0:
+            # We could just let the negative values get truncated to zero, but
+            # empirically the results seem better when we shift all parameters up.
+            a[:] -= min_val
+
+    return a
+
+
+# This approach is unsatisfactory as we should minimise the L2 norm
+# rather than the L1 norm.
+#
+# def _solve_pulp_l1(m: dok_matrix, b: np.ndarray) -> np.ndarray:
+#     """
+#     Find the best 'a' for `b = m a` subject to `a[i] >= 0`.
+#
+#     Uses pulp LpProblem to minimise the L1 norm of `a`.
+#
+#     This method will only work if there is an exact solution.
+#     If not, then we call _solve_sam as a fallback.
+#     """
+#     import pulp
+#
+#     a_size = m.shape[1]
+#     b_size = b.shape[0]
+#
+#     prob = pulp.LpProblem('solver', pulp.LpMinimize)
+#     x = [pulp.LpVariable(f'x{i}', lowBound=0) for i in range(a_size)]
+#
+#     # The objective: minimise the L1 norm of x.
+#     # The sum(x) is the L1 norm because each element of x is constrained >= 0.
+#     prob.setObjective(pulp.lpSum(x))
+#
+#     # The constraints
+#     constraints = [pulp.LpAffineExpression() for _ in range(b_size)]
+#     for row, col in m.keys():
+#         constraints[row].addterm(x[col], 1)
+#     for c, b_i in zip(constraints, b):
+#         prob.addConstraint(c == b_i)
+#
+#     _PULP_TIMEOUT = 60  # seconds
+#     status = prob.solve(pulp.PULP_CBC_CMD(msg=False, timeLimit=_PULP_TIMEOUT))
+#
+#     if status == pulp.LpStatusOptimal:
+#         return np.array([pulp.value(x_var) for x_var in x])
+#     else:
+#         return _solve_sam(m, b)
+
+
+def _sum_out_unneeded_rvs(crosstab: CrossTable, rvs: Sequence[RandomVariable]) -> CrossTable:
+    """
+    Project the given cross-table as needed to ensure all random
+    variables in the result are in `rvs`.
+    """
+    available_rvs = set(crosstab.rvs)
+    project_rvs = available_rvs.intersection(rvs)
+    if len(project_rvs) == len(available_rvs):
+        # No projection is required
+        return crosstab
+    else:
+        return crosstab.project(list(project_rvs))
+
+
+def _make_matrix(
+        crosstabs: Sequence[CrossTable],
+        rvs: Sequence[RandomVariable],
+) -> Tuple[dok_matrix, np.ndarray, Sequence[Instance]]:
+    """
+    Create the `m` matrix and `b` vector for solving `b = m a`.
+
+    Args:
+        crosstabs: a collection of cross-tables to coalesce.
+        rvs: the random variables that cross-tables will be projected on to.
+
+    Returns:
+        the tuple (m, b, a_keys) where
+        'm' is a sparse matrix,
+        'b' is a numpy array of crosstab probabilities,
+        'a_keys' are the keys for the solution probabilities.
+    """
+
+    # Sum out any unneeded random variables
+    crosstabs: Sequence[CrossTable] = [
+        _sum_out_unneeded_rvs(crosstab, rvs)
+        for crosstab in crosstabs
+    ]
+
+    m_cols: Dict[Instance, _MCol] = {}
+    b_list: List[float] = []
+    a_keys: List[Instance] = []
+
+    rv_index: Dict[RandomVariable, int] = {rv: i for i, rv in enumerate(rvs)}
+
+    # instance_template[i] is a list of the possible states of rvs[i]
+    instance_template: List[List[int]] = [list(range(len(rv))) for rv in enumerate(rvs)]
+
+    for crosstab in crosstabs:
+
+        # get `to_rv` such that crosstab.rvs[i] = rvs[to_rv[i]]
+        to_rv = [rv_index.get(rv) for rv in crosstab.rvs]
+
+        # Make instance_options which is a clone of instance_template but with
+        # a singleton list replacing the rvs that this crosstab has.
+        # For now the state in each singleton is set to -1, however, later
+        # they will be set to the actual states of instances in the current crosstab.
+        instance_options = list(instance_template)
+        for i in to_rv:
+            instance_options[i] = [-1]
+
+        total = crosstab.total_weight()
+        for crosstab_instance, weight in crosstab.items():
+
+            # Work out what instances get summed to create the crosstab_instance weight.
+            # This just overrides the singleton states of `instance_options` with the
+            # actual state of the crosstab_instance.
+            for state, i in zip(crosstab_instance, to_rv):
+                instance_options[i][0] = state
+
+            # Grow the b list with our instance probability
+            b_i = len(b_list)
+            b_list.append(weight / total)
+
+            # Iterate over all states of `rvs` that matches the current crosstab_instance
+            # recording `b_i` in the column for those matching instances.
+            for instance in combos(instance_options):
+                m_col = m_cols.get(instance)
+                if m_col is None:
+                    m_cols[instance] = _MCol(instance, len(a_keys), [b_i])
+                    a_keys.append(instance)
+                else:
+                    m_col.col.append(b_i)
+
+    # Construct the m matrix from m_cols
+    m = dok_matrix((len(b_list), len(a_keys)), dtype=np.double)
+    for m_col in m_cols.values():
+        j = m_col.column_index
+        for i in m_col.col:
+            m[i, j] = 1
+
+    # Construct the b vector
+    b = np.array(b_list, dtype=np.double)
+
+    return m, b, a_keys
+
+
+@dataclass
+class _MCol:
+    key: Instance
+    column_index: int
+    col: List[int]
+
+
+@dataclass
+class _SM:
+    split: float
+    a: float
+
+    def diff(self) -> float:
+        return self.split - self.a
+
+
+class _SAM:
+    """
+    Split and Mean method for finding 'a'
+    in b = m a
+    subject to a[i] >= 0.
+
+    Assumes all elements of `m` are either zero or one.
+    """
+
+    def __init__(self, m: dok_matrix, b: NDArrayFloat64, use_lsqr: bool = True):
+        """
+        Allocate the memory required for a SAM solver.
+
+        Args:
+            m: the summation matrix
+            b: the vector of resulting probabilities
+            use_lsqr: whether to use LSQR or not to initialise the solution.
+        """
+        # Replicate the sparse m matrix, as a list of lists of _SM objects,
+        # where we have both row major and column major representations.
+        a_size: int = m.shape[1]
+        b_size: int = m.shape[0]
+        a_idx: List[List[_SM]] = [[] for _ in range(a_size)]
+        b_idx: List[List[_SM]] = [[] for _ in range(b_size)]
+        for (i, j), m_val in m.items():
+            if m_val != 0:
+                sm = _SM(0, 0)
+                a_idx[j].append(sm)
+                b_idx[i].append(sm)
+
+        self._a: NDArrayFloat64 = np.zeros(a_size, dtype=np.double)
+        self._b: NDArrayFloat64 = b
+        self._m: dok_matrix = m
+        self._a_idx: List[List[_SM]] = a_idx
+        self._b_idx: List[List[_SM]] = b_idx
+        self._use_lsqr = use_lsqr
+
+    def solve(self, max_iterations: int, tolerance: float, change_tolerance: float) -> Tuple[np.ndarray, float]:
+        """
+        Initialize split values then iterate (mean step, split step).
+
+        Args:
+            max_iterations: maximum number of iterations.
+            tolerance: terminate iterations if error <= tolerance.
+            change_tolerance: terminate iterations if change in error <= change_tolerance.
+
+        Returns:
+            tuple ('a', 'error') where 'a' is the current solution after this step
+            and 'error' is the sum of absolute errors.
+        """
+        self._initialize_split_values()
+        iteration = 0
+        prev_error = 0
+        while True:
+            iteration += 1
+            a, error = self._mean_step()
+            if error <= tolerance or abs(error - prev_error) <= change_tolerance or iteration >= max_iterations:
+                return a, error
+            prev_error = error
+            self._split_step()
+
+    def _initialize_split_values(self):
+        """
+        Take each 'b' value and split it across its SM cells.
+
+        If 'self._use_lsqr' is True then the split is based on a solution
+        using scipy lsqr, otherwise the split is even for each 'b' value.
+        """
+        for b_val, b_list in zip(self._b, self._b_idx):
+            len_b_list = len(b_list)
+            if len_b_list > 0:
+                split_val = b_val / len_b_list
+                for sm in b_list:
+                    sm.split = split_val
+
+        if self._use_lsqr:
+            a = _solve_lsqr(self._m, self._b)
+            assert len(a) == len(self._a)
+            for a_val, a_list in zip(a, self._a_idx):
+                for sm in a_list:
+                    sm.a = a_val
+            self._split_step()
+
+    def _mean_step(self) -> Tuple[np.ndarray, float]:
+        """
+        Take the current split values to determine the 'a' values
+        as the mean across relevant SM cells.
+
+        Assumes the previous step was either 'initialize_split_values'
+        or a 'split step'.
+
+        Returns:
+            tuple ('a', 'error') where 'a' is the current solution after this step
+            and 'error' is the sum of absolute errors.
+        """
+        error = 0.0
+        a = self._a
+        for i, a_list in enumerate(self._a_idx):
+            sum_val = sum(sm.split for sm in a_list)
+            a_val = sum_val / len(a_list)
+            a[i] = a_val
+            for sm in a_list:
+                sm.a = a_val
+                error += abs(sm.diff())
+        return a, error
+
+    def _split_step(self):
+        """
+        Take the difference between the split 'b' values and current 'a' values to
+        redistribute the split 'b' values.
+
+        Assumes the previous step was a 'mean step'.
+        """
+        for b_val, b_list in zip(self._b, self._b_idx):
+            if len(b_list) <= 1:
+                # Too small to split
+                continue
+            pos = 0
+            neg = 0
+            for sm in b_list:
+                diff = sm.diff()
+                if diff >= 0:
+                    pos += diff
+                else:
+                    neg -= diff
+            mass = min(pos, neg)
+            if mass == 0:
+                # No mass to redistribute
+                continue
+            pos = mass / pos
+            neg = mass / neg
+            for sm in b_list:
+                diff = sm.diff()
+                if diff >= 0:
+                    mass = diff * pos
+                else:
+                    mass = diff * neg
+                sm.split -= mass

ck/learning/model_from_cross_tables.py

@@ -0,0 +1,242 @@
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from itertools import chain
+from typing import Iterable, List, Tuple, Dict, Sequence, Set, Optional
+
+from ck.dataset.cross_table import CrossTable
+from ck.learning.coalesce_cross_tables import coalesce_cross_tables
+from ck.learning.parameters import make_factors, ParameterValues
+from ck.learning.train_generative_bn import cpt_and_parent_sums_from_crosstab
+from ck.pgm import PGM, RandomVariable
+from ck.utils.map_list import MapList
+
+
+def model_from_cross_tables(pgm: PGM, cross_tables: Iterable[CrossTable]) -> None:
+    """
+    Make best efforts to construct a Bayesian network model given only the
+    evidence from the supplied cross-tables.
+
+    This function calls `get_cpts` which provides parameters to define
+    a Bayesian network model. These are then applied to the PGM using
+    `make_factors`.
+
+    Args:
+        pgm: the PGM to add factors and potential function to.
+        cross_tables: available cross-tables to build a model from.
+
+    Raises:
+        ValueError: If `pgm` has any existing factors.
+    """
+    if len(pgm.factors) > 0:
+        raise ValueError('the given PGM should have no factors')
+    cpts: List[CrossTable] = get_cpts(
+        rvs=pgm.rvs,
+        cross_tables=cross_tables,
+    )
+    make_factors(pgm, cpts)
+
+
+def get_cpts(rvs: Sequence[RandomVariable], cross_tables: Iterable[CrossTable]) -> ParameterValues:
+    """
+    Make best efforts to define a Bayesian network model given only the
+    evidence from the supplied cross-tables.
+
+    This function infers CPTs for `rvs` using the given `cross_tables`.
+
+    For any two cross-tables `x` and `y` in `cross_tables`, with common random
+    variables `rvs` then this function assumes `x.project(rvs) == y.project(rvs)`.
+    If this condition does not hold, then best efforts will still be made to
+    define a Bayesian network model, however, the resulting parameter values
+    may be suboptimal.
+
+    Args:
+        rvs: the random variables to define a network structure over.
+        cross_tables: available cross-tables to build a model from.
+
+    Returns:
+        ParameterValues object as a list of CPTs, each CPT can be used to create
+        a new factor in the given PGM to make a Bayesian network.
+    """
+    # Stabilise the given crosstables
+    cross_tables: Tuple[CrossTable, ...] = tuple(cross_tables)
+
+    # Heuristically determine an ordering over the random variables
+    # which will be used to form the BN structure.
+    rv_order: Dict[RandomVariable, int] = _get_rv_order(cross_tables)
+
+    # Make an empty model factor for each random variable.
+    model_factors: List[_ModelFactor] = [_ModelFactor(rv) for rv in rvs]
+
+    # Define a Bayesian network structure.
+    # Allocate each crosstab to exactly one random variable, the
+    # one with the highest rank (i.e. the child).
+    for crosstab in cross_tables:
+        if len(crosstab.rvs) == 0:
+            continue
+        sorted_rvs: List[RandomVariable] = sorted(crosstab.rvs, key=(lambda _rv: rv_order[_rv]), reverse=True)
+        child: RandomVariable = sorted_rvs[0]
+        parents: List[RandomVariable] = sorted_rvs[1:]
+        model_factor: _ModelFactor = model_factors[child.idx]
+        model_factor.parent_rvs.update(parents)
+        model_factor.cross_tables.append(crosstab)
+
+    # Link child factors.
+    # When defining a factor, we need to define the child factors first.
+    for model_factor in model_factors:
+        for parent_rv in model_factor.parent_rvs:
+            model_factors[parent_rv.idx].child_factors.append(model_factor)
+
+    # Make the factors, depth first.
+    done: Set[int] = set()
+    for model_factor in model_factors:
+        _infer_cpt(model_factor, done)
+
+    # Return the CPTs that define the structure
+    return [model_factor.cpt for model_factor in model_factors]
+
+
+def _infer_cpt(model_factor: _ModelFactor, done: Set[int]) -> None:
+    """
+    Depth-first recursively infer the model factors as CPTs.
+    This sets `model_factor.cpt` and `model_factor.parent_rvs` for
+    the given `model_factor` and its children.
+    """
+    # Only visit a model factor once.
+    child_rv: RandomVariable = model_factor.child_rv
+    if child_rv.idx in done:
+        return
+    done.add(child_rv.idx)
+
+    # Recursively visit the child factors
+    for child_model_factor in model_factor.child_factors:
+        _infer_cpt(child_model_factor, done)
+
+    # Get all relevant cross-tables to set the parameters
+    crosstabs: Sequence[CrossTable] = model_factor.cross_tables
+    child_crosstabs: Sequence[CrossTable] = [
+        child_model_factor.parent_sums
+        for child_model_factor in model_factor.child_factors
+    ]
+
+    # Get the parameters
+    rvs = [child_rv] + list(model_factor.parent_rvs)
+    crosstab: CrossTable = _infer_parameters(rvs, crosstabs, child_crosstabs)
+    cpt, parent_sums = cpt_and_parent_sums_from_crosstab(crosstab)
+    model_factor.cpt = cpt
+    model_factor.parent_sums = parent_sums
+
+
+def _infer_parameters(
+        rvs: Sequence[RandomVariable],
+        crosstabs: Sequence[CrossTable],
+        child_crosstabs: Sequence[CrossTable],
+) -> CrossTable:
+    """
+    Make best efforts to infer a probability distribution over the give random variables,
+    with evidence from the given cross-tables.
+
+    Returns:
+        a CrossTable representing the inferred probability distribution
+        (not normalised to sum to one).
+
+    Assumes:
+        `rvs` has no duplicates.
+    """
+
+    # Project crosstables onto rvs, splitting them into complete and partial coverage of `rvs`.
+    # Completely covering cross-tables will be accumulated into `complete_crosstab` while others
+    # will be appended to partial_crosstabs.
+    complete_crosstab: Optional[CrossTable] = None
+    partial_crosstabs: List[CrossTable] = []
+    for available_crosstab in chain(crosstabs, child_crosstabs):
+        available_rvs: Set[RandomVariable] = set(available_crosstab.rvs)
+        if available_rvs.issuperset(rvs):
+            to_add: CrossTable = available_crosstab.project(rvs)
+            if complete_crosstab is None:
+                complete_crosstab = to_add
+            else:
+                complete_crosstab.add_all(to_add.project(rvs).items())
+        else:
+            partial_crosstabs.append(available_crosstab)
+
+    if complete_crosstab is not None:
+        # A direct solution was found.
+        # Ignore any partially covering cross-tables.
+        return complete_crosstab
+
+    # If there are no cross-tables available, the result is empty
+    if len(partial_crosstabs) == 0:
+        return CrossTable(rvs)
+
+    # There were no cross-tables that completely cover the given random variables.
+    # The following algorithm makes best attempts to coalesce a collection of
+    # partially covering cross-tables.
+
+    return coalesce_cross_tables(partial_crosstabs, rvs)
+
+
+def _crostab_str(crosstab: CrossTable) -> str:
+    return '(' + ', '.join(repr(rv.name) for rv in crosstab.rvs) + ')'
+
+
+def _get_rv_order(cross_tables: Sequence[CrossTable]) -> Dict[RandomVariable, int]:
+    """
+    Determine an order over the given random variables.
+    Returns a map from rv to its rank in the order.
+    """
+    child_parent_map: MapList[RandomVariable, RandomVariable] = MapList()
+    for crosstab in cross_tables:
+        rvs = crosstab.rvs
+        for i in range(len(rvs)):
+            child = rvs[i]
+            parents = rvs[i + 1:]
+            child_parent_map.extend(child, parents)
+    order: Dict[RandomVariable, int] = {}
+    seen: Set[RandomVariable] = set()
+    for child in child_parent_map.keys():
+        _get_rv_order_r(child, child_parent_map, seen, order)
+    return order
+
+
+def _get_rv_order_r(
+        child: RandomVariable,
+        child_parent_map: MapList[RandomVariable, RandomVariable],
+        seen: Set[RandomVariable],
+        order: Dict[RandomVariable, int],
+):
+    if child not in seen:
+        seen.add(child)
+        parents = child_parent_map[child]
+        for parent in parents:
+            _get_rv_order_r(parent, child_parent_map, seen, order)
+        order[child] = len(order)
+
+
+@dataclass
+class _ModelFactor:
+    """
+    A collection of model factors defines a PGM structure,
+    each model factor representing a needed PGM factor.
+
+    Associated with a model factor is a list of crosstabs
+    that whose rvs overlap with the rvs of the model factor.
+    """
+
+    child_rv: RandomVariable
+    parent_rvs: set[RandomVariable] = field(default_factory=set)
+    cross_tables: List[CrossTable] = field(default_factory=list)
+    child_factors: List[_ModelFactor] = field(default_factory=list)
+
+    # These are set once the parameter values are inferred...
+    cpt: Optional[CrossTable] = None
+    parent_sums: Optional[CrossTable] = None
+
+    def dump(self, prefix=''):
+        """
+        For debugging.
+        """
+        print(f'{prefix}child:', self.child_rv)
+        print(f'{prefix}parents:', '{}' if not self.parent_rvs else self.parent_rvs)
+        for cross_table in self.cross_tables:
+            print(f'{prefix}cross-table:', *cross_table.rvs)