compiled-knowledge 4.0.0a24__cp313-cp313-win_amd64.whl → 4.1.0a1__cp313-cp313-win_amd64.whl

ck/dataset/dataset.py

@@ -0,0 +1,577 @@
+from __future__ import annotations
+
+from typing import Sequence, Optional, Dict, Iterable, Tuple
+
+import numpy as np
+
+from ck.pgm import RandomVariable, State
+from ck.utils.np_extras import NDArray, DTypeStates, dtype_for_number_of_states
+
+
+class Dataset:
+    """
+    A dataset has instances (rows) for zero or more random variables.
+    Each instance has a weight, which is notionally one.
+    Weights of instances should be non-negative, and are normally positive.
+    """
+
+    def __init__(
+            self,
+            weights: Optional[NDArray | Sequence],
+            length: Optional[int],
+    ):
+        # Infer the length of the dataset.
+        if length is not None:
+            self._length: int = length
+        else:
+            self._length: int = len(weights)
+
+        # Set no random variables
+        self._rvs: Tuple[RandomVariable, ...] = ()
+
+        # Set the weights array, and confirm its shape
+        self._weights: NDArray
+        if weights is None:
+            weights = np.ones(self._length)
+        elif not isinstance(weights, np.ndarray):
+            weights = np.array(weights)
+        expected_shape = (self._length,)
+        if weights.shape != expected_shape:
+            raise ValueError(f'weights expected shape {expected_shape}, got {weights.shape}')
+
+        self._weights = weights
+
+    def __len__(self) -> int:
+        """
+        How many instances in the dataset.
+        """
+        return self._length
+
+    @property
+    def rvs(self) -> Sequence[RandomVariable]:
+        """
+        Return the random variables covered by this dataset.
+        """
+        return self._rvs
+
+    @property
+    def weights(self) -> NDArray:
+        """
+        Get the instance weights.
+        The notional weight of an instance is 1.
+        The index into the returned array is the instance index.
+
+        Returns:
+            A 1D array of random variable states, with shape = `(len(self), )`.
+        """
+        return self._weights
+
+    def total_weight(self) -> float:
+        """
+        Calculate the total weight of this dataset.
+        """
+        return self._weights.sum().item()
+
+    def _add_rv(self, rv: RandomVariable) -> None:
+        """
+        Add a random variable to self.rvs.
+        """
+        self._rvs += (rv,)
+
+    def _remove_rv(self, rv: RandomVariable) -> None:
+        """
+        Remove a random variable from self.rvs.
+        """
+        rvs = self._rvs
+        i: int = self._rvs.index(rv)
+        self._rvs = rvs[:i] + rvs[i+1:]
+
+
+class HardDataset(Dataset):
+    """
+    A hard dataset is a dataset where for each instance (row) and each random variable,
+    there is a state for that random variable (a state is represented as a state index).
+    Each instance has a weight, which is notionally one.
+    """
+
+    @staticmethod
+    def from_soft_dataset(
+            soft_dataset: SoftDataset,
+            adjust_instance_weights: bool = True,
+    ) -> HardDataset:
+        """
+        Create a hard dataset from a soft dataset by repeated application
+        of `SoftDataset.add_rv_from_state_weights`.
+
+        The instance weights of the returned dataset will be a copy
+        of the instance weights of the soft dataset.
+
+        Args:
+            soft_dataset: The soft dataset providing random variables,
+                their states, and instance weights.
+            adjust_instance_weights: If `True` (default), then the instance weights will be
+                adjusted according to sum of state weights for each instance. That is, if
+                the sum is not one for some instance, then the weight of that instance will
+                be adjusted.
+
+        Returns:
+            A `HardDataset` instance.
+        """
+        dataset = HardDataset(weights=soft_dataset.weights.copy())
+        for rv in soft_dataset.rvs:
+            dataset.add_rv_from_state_weights(rv, soft_dataset.state_weights(rv), adjust_instance_weights)
+        return dataset
+
+    def __init__(
+            self,
+            data: Iterable[Tuple[RandomVariable, NDArray | Sequence[int]]] = (),
+            weights: Optional[NDArray | Sequence] = None,
+            length: Optional[int] = None,
+    ):
+        """
+        Create a hard dataset.
+
+        Args:
+            data: optional iterable of (random variable, state idxs), passed
+                to `self.add_rv_from_state_idxs`.
+            weights: optional array of instance weights.
+            length: optional length of the dataset, if omitted, the length is inferred.
+        """
+        self._data: Dict[RandomVariable, NDArray] = {}
+
+        # Initialise super by either weights, length or first data item.
+        super_initialised: bool = False
+        if weights is not None or length is not None:
+            super().__init__(weights, length)
+            super_initialised = True
+
+        for rv, states in data:
+            if not super_initialised:
+                super().__init__(weights, len(states))
+                super_initialised = True
+            self.add_rv_from_state_idxs(rv, states)
+
+        if not super_initialised:
+            super().__init__(weights, 0)
+
+    def states(self, rv: RandomVariable) -> NDArray:
+        """
+        Get the state values (by state index) for one random variable.
+        The index into the returned array is the instance index.
+
+        Returns:
+            A 1D array of random variable states, with shape = `(len(self), )`.
+
+        Raises:
+            KeyError: If the random variable is not in the dataset.
+        """
+        return self._data[rv]
+
+    def add_rv(self, rv: RandomVariable) -> NDArray:
+        """
+        Add a random variable to the dataset, allocating and returning
+        the state indices for the random variable.
+
+        Args:
+            rv: The random variable to add.
+
+        Returns:
+            A 1D array of random variable states, with shape = `(len(self), )`, initialised to zero.
+
+        Raises:
+            ValueError: If the random variable is already in the dataset.
+        """
+        dtype: DTypeStates = dtype_for_number_of_states(len(rv))
+        rv_data = np.zeros(len(self), dtype=dtype)
+        return self.add_rv_from_state_idxs(rv, rv_data)
+
+    def remove_rv(self, rv: RandomVariable) -> None:
+        """
+        Remove a random variable from the dataset.
+
+        Args:
+            rv: The random variable to remove.
+
+        Raises:
+            KeyError: If the random variable is not in the dataset.
+        """
+        del self._data[rv]
+        self._remove_rv(rv)
+
+    def add_rv_from_state_idxs(self, rv: RandomVariable, state_idxs: NDArray | Sequence[int]) -> NDArray:
+        """
+        Add a random variable to the dataset.
+
+        The dataset will directly reference the given `states` array.
+
+        Args:
+            rv: The random variable to add.
+            state_idxs: An 1D array of state indexes to add, with shape = `(len(self),)`.
+                Each element `state` should be `0 <= state < len(rv)`.
+
+        Returns:
+            A 1D array of random variable states, with shape = `(len(self), )`.
+
+        Raises:
+            ValueError: If the random variable is already in the dataset.
+        """
+        if rv in self._data.keys():
+            raise ValueError(f'data for {rv} already exists in the dataset')
+
+        if isinstance(state_idxs, np.ndarray):
+            expected_shape = (self._length,)
+            if state_idxs.shape == expected_shape:
+                rv_data = state_idxs
+            else:
+                raise ValueError(f'data for {rv} expected shape {expected_shape}, got {state_idxs.shape}')
+        else:
+            dtype: DTypeStates = dtype_for_number_of_states(len(rv))
+            if len(state_idxs) != self._length:
+                raise ValueError(f'data for {rv} expected length {self._length}, got {len(state_idxs)}')
+            rv_data = np.array(state_idxs, dtype=dtype)
+
+        self._data[rv] = rv_data
+        self._add_rv(rv)
+        return rv_data
+
+    def add_rv_from_states(self, rv: RandomVariable, states: Sequence[State]) -> NDArray:
+        """
+        Add a random variable to the dataset.
+
+        The dataset will allocate and populate a states array containing state indexes.
+        This will call `rv.state_idx(state)` for each state in `states`.
+
+        Args:
+            rv: The random variable to add.
+            states: An 1D array of state to add, with `len(states)` = `len(self)`.
+                Each element `state` should be in `rv.states`.
+
+        Returns:
+            A 1D array of random variable states, with shape = `(len(self), )`.
+
+        Raises:
+            ValueError: If the random variable is already in the dataset.
+        """
+        dtype: DTypeStates = dtype_for_number_of_states(len(rv))
+        rv_data = np.fromiter(
+            iter=(
+                rv.state_idx(state)
+                for state in states
+            ),
+            dtype=dtype,
+            count=len(states)
+        )
+        return self.add_rv_from_state_idxs(rv, rv_data)
+
+    def add_rv_from_state_weights(
+            self,
+            rv: RandomVariable,
+            state_weights: NDArray,
+            adjust_instance_weights: bool = True,
+    ) -> NDArray:
+        """
+        Add a random variable to the dataset.
+
+        The dataset will allocate and populate a states array containing state indexes.
+        For each instance, the state with the highest weight will be taken to be the
+        state of the random variable, with ties broken arbitrarily.
+
+        Args:
+            rv: The random variable to add.
+            state_weights: An 2D array of state weights, with shape = `(len(self), len(rv))`.
+                Each element `state` should be in `rv.states`.
+            adjust_instance_weights: If `True` (default), then the instance weights will be
+                adjusted according to sum of state weights for each instance. That is, if
+                the sum is not one for some instance, then the weight of that instance will
+                be adjusted.
+
+        Returns:
+            A 1D array of random variable states, with shape = `(len(self), )`.
+
+        Raises:
+            ValueError: If the random variable is already in the dataset.
+        """
+        expected_shape = (self._length, len(rv))
+        if state_weights.shape != expected_shape:
+            raise ValueError(f'data for {rv} expected shape {expected_shape}, got {state_weights.shape}')
+
+        dtype: DTypeStates = dtype_for_number_of_states(len(rv))
+        rv_data = np.fromiter(
+            iter=(
+                np.argmax(row)
+                for row in state_weights
+            ),
+            dtype=dtype,
+            count=self._length
+        )
+
+        if adjust_instance_weights:
+            row: NDArray
+            for i, row in enumerate(state_weights):
+                self._weights[i] *= row.sum()
+
+        return self.add_rv_from_state_idxs(rv, rv_data)
+
+    def dump(self, *, show_rvs: bool = True, show_weights: bool = True, as_states: bool = False) -> None:
+        """
+        Dump the dataset in a human-readable format.
+        If as_states is true, then instance states are dumped instead of just state indexes.
+
+        Args:
+            show_rvs: If `True`, the random variables are dumped.
+            show_weights: If `True`, the instance weights are dumped.
+            as_states: If `True`, the states are dumped Instead of just state indexes.
+        """
+        if show_rvs:
+            rvs = ', '.join(str(rv) for rv in self.rvs)
+            print(f'rvs: [{rvs}]')
+        print(f'instances ({len(self)}, with total weight {self.total_weight()}):')
+        cols = [self.states(rv) for rv in self.rvs]
+        for instance, weight in zip(zip(*cols), self.weights):
+            if as_states:
+                instance_str = ', '.join(repr(rv.states[idx]) for idx, rv in zip(instance, self.rvs))
+            else:
+                instance_str = ', '.join(str(idx) for idx in instance)
+            if show_weights:
+                print(f'({instance_str}) * {weight}')
+            else:
+                print(f'({instance_str})')
+
+
+class SoftDataset(Dataset):
+    """
+    A soft dataset is a dataset where for each instance (row) and each random variable,
+    there is a distribution over the states of that random variable. That is,
+    for each instance, for each indicator, there is a weight. Additionally,
+    each instance has a weight.
+
+    Weights of random variable states are expected to be non-negative.
+    Notionally, the sum of weights for an instance and random variable is one.
+    """
+
+    @staticmethod
+    def from_hard_dataset(hard_dataset: HardDataset) -> SoftDataset:
+        """
+        Create a soft dataset from a hard dataset by repeated application
+        of `SoftDataset.add_rv_from_state_idxs`.
+
+        The instance weights of the returned dataset will be a copy
+        of the instance weights of the hard dataset.
+
+        Args:
+            hard_dataset: The hard dataset providing random variables,
+                their states, and instance weights.
+
+        Returns:
+            A `SoftDataset` instance.
+        """
+        dataset = SoftDataset(weights=hard_dataset.weights.copy())
+        for rv in hard_dataset.rvs:
+            dataset.add_rv_from_state_idxs(rv, hard_dataset.states(rv))
+        return dataset
+
+    def __init__(
+            self,
+            data: Iterable[Tuple[RandomVariable, NDArray]] = (),
+            weights: Optional[NDArray | Sequence] = None,
+            length: Optional[int] = None,
+    ):
+        """
+        Create a soft dataset.
+
+        Args:
+            data: optional iterable of (random variable, state weights), passed
+                to `self.add_rv_from_state_weights`.
+            weights: optional array of instance weights.
+            length: optional length of the dataset, if omitted, the length is inferred.
+        """
+        self._data: Dict[RandomVariable, NDArray] = {}
+
+        # Initialise super by either weights, length or first data item.
+        super_initialised: bool = False
+        if weights is not None or length is not None:
+            super().__init__(weights, length)
+            super_initialised = True
+
+        for rv, states_weights in data:
+            if not super_initialised:
+                super().__init__(weights, len(states_weights))
+                super_initialised = True
+            self.add_rv_from_state_weights(rv, states_weights)
+
+        if not super_initialised:
+            super().__init__(weights, 0)
+
+    def normalise(self, check_negative_instance: bool = True) -> None:
+        """
+        Adjust weights (for states and instances) so that the sum of state weights
+        for any random variable is 1 (or zero).
+
+        This performs an in-place modification.
+
+        If an instance weight is zero then all state weights for that instance will be zero.
+        If the state weights of an instance for any random variable sum to zero, then
+        that instance weight will be zero.
+
+        All other state weights of an instance for each random variable will sum to one.
+
+        Args:
+            check_negative_instance: if true (the default),then a RuntimeError is
+                raised if a negative instance weight is encountered.
+
+        Raises:
+            RuntimeError: if `check_negative_instance` is true and a negative
+                instance weight is encountered.
+        """
+        state_weights: NDArray
+        i: int
+
+        weights: NDArray = self.weights
+        for i in range(self._length):
+            for state_weights in self._data.values():
+                weight_sum = state_weights[i].sum()
+                if weight_sum == 0:
+                    weights[i] = 0
+                elif weight_sum != 1:
+                    state_weights[i] /= weight_sum
+                    weights[i] *= weight_sum
+            instance_weight = weights[i]
+            if instance_weight == 0:
+                for state_weights in self._data.values():
+                    state_weights[i, :] = 0
+            elif check_negative_instance and instance_weight < 0:
+                raise RuntimeError(f'negative instance weight: {i}')
+
+    def state_weights(self, rv: RandomVariable) -> NDArray:
+        """
+        Get the state weights for one random variable.
+        The first index into the returned array is the instance index.
+        The second index into the returned array is the state index.
+
+        Returns:
+            A 2D array of random variable states, with shape = `(len(self), len(rv))`.
+
+        Raises:
+            KeyError: If the random variable is not in the dataset.
+        """
+        return self._data[rv]
+
+    def add_rv(self, rv: RandomVariable) -> NDArray:
+        """
+        Add a random variable to the dataset, allocating and returning
+        the state indices for the random variable.
+
+        Args:
+            rv: The random variable to add.
+
+        Returns:
+            A 2D array of random variable states, with shape = `(len(self), len(rv))`,
+            initialised to zero.
+
+        Raises:
+            ValueError: If the random variable is already in the dataset.
+        """
+        rv_data = np.zeros((len(self), len(rv)), dtype=np.float64)
+        return self.add_rv_from_state_weights(rv, rv_data)
+
+    def remove_rv(self, rv: RandomVariable) -> None:
+        """
+        Remove a random variable from the dataset.
+
+        Args:
+            rv: The random variable to remove.
+
+        Raises:
+            KeyError: If the random variable is not in the dataset.
+        """
+        del self._data[rv]
+        self._remove_rv(rv)
+
+    def add_rv_from_state_weights(self, rv: RandomVariable, state_weights: NDArray) -> NDArray:
+        """
+        Add a random variable to the dataset.
+
+        The dataset will directly reference the given `states` array.
+
+        Args:
+            rv: The random variable to add.
+            state_weights: A 2D array of state weights, with shape = `(len(self), len(rv))`.
+
+        Raises:
+            ValueError: If the random variable is already in the dataset.
+        """
+        if rv in self._data.keys():
+            raise ValueError(f'data for {rv} already exists in the dataset')
+
+        expected_shape = (self._length, len(rv))
+        if state_weights.shape == expected_shape:
+            rv_data = state_weights
+        else:
+            raise ValueError(f'data for {rv} expected shape {expected_shape}, got {state_weights.shape}')
+
+        self._data[rv] = rv_data
+        self._add_rv(rv)
+        return rv_data
+
+    def add_rv_from_state_idxs(self, rv: RandomVariable, state_idxs: NDArray | Sequence[int]) -> NDArray:
+        """
+        Add a random variable to the dataset.
+
+        The dataset will directly reference the given `states` array.
+
+        Args:
+            rv: The random variable to add.
+            state_idxs: An 1D array of state indexes to add, with shape = `(len(self),)`.
+                Each element `state` should be `0 <= state < len(rv)`.
+
+        Raises:
+            ValueError: If the random variable is already in the dataset.
+        """
+        rv_data = np.zeros((len(state_idxs), len(rv)), dtype=np.float64)
+        for i, state_idx in enumerate(state_idxs):
+            rv_data[i, state_idx] = 1
+
+        return self.add_rv_from_state_weights(rv, rv_data)
+
+    def add_rv_from_states(self, rv: RandomVariable, states: Sequence[State]) -> NDArray:
+        """
+        Add a random variable to the dataset.
+
+        The dataset will allocate and populate a states array containing state indexes.
+        This will call `rv.state_idx(state)` for each state in `states`.
+
+        Args:
+            rv: The random variable to add.
+            states: An 1D array of state to add, with `len(states)` = `len(self)`.
+                Each element `state` should be in `rv.states`.
+
+        Raises:
+            ValueError: If the random variable is already in the dataset.
+        """
+        rv_data = np.zeros((len(states), len(rv)), dtype=np.float64)
+        for i, state in enumerate(states):
+            state_idx = rv.state_idx(state)
+            rv_data[i, state_idx] = 1
+
+        return self.add_rv_from_state_weights(rv, rv_data)
+
+    def dump(self, *, show_rvs: bool = True, show_weights: bool = True) -> None:
+        """
+        Dump the dataset in a human-readable format.
+        If as_states is true, then instance states are dumped instead of just state indexes.
+
+        Args:
+            show_rvs: If `True`, the random variables are dumped.
+            show_weights: If `True`, the instance weights are dumped.
+        """
+        if show_rvs:
+            rvs = ', '.join(str(rv) for rv in self.rvs)
+            print(f'rvs: [{rvs}]')
+        print(f'instances ({len(self)}, with total weight {self.total_weight()}):')
+        cols = [self.state_weights(rv) for rv in self.rvs]
+        for instance, weight in zip(zip(*cols), self.weights):
+            instance_str = ', '.join(str(state_weights) for state_weights in instance)
+            if show_weights:
+                print(f'({instance_str}) * {weight}')
+            else:
+                print(f'({instance_str})')