ailia-models/code/pyannote_audio_utils/audio/utils/powerset.py

# MIT License
#
# Copyright (c) 2023- CNRS
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.

# AUTHORS
# Hervé BREDIN - https://herve.niderb.fr
# Alexis PLAQUET

from functools import cached_property
from itertools import combinations

import scipy.special
import numpy as np

class Powerset():
    """Powerset to multilabel conversion, and back.

    Parameters
    ----------
    num_classes : int
        Number of regular classes.
    max_set_size : int
        Maximum number of classes in each set.
    """

    def __init__(self, num_classes: int, max_set_size: int):
        super().__init__()
        self.num_classes = num_classes
        self.max_set_size = max_set_size
        self.mapping = self.build_mapping()
        self.cardinality = self.build_cardinality()
        
        

    @cached_property
    def num_powerset_classes(self) -> int:
        # compute number of subsets of size at most "max_set_size"
        # e.g. with num_classes = 3 and max_set_size = 2:
        # {}, {0}, {1}, {2}, {0, 1}, {0, 2}, {1, 2}
        return int(
            sum(
                scipy.special.binom(self.num_classes, i)
                for i in range(0, self.max_set_size + 1)
            )
        )

    def build_mapping(self) -> np.ndarray:
        """Compute powerset to regular mapping

        Returns
        -------
        mapping : (num_powerset_classes, num_classes) torch.Tensor
            mapping[i, j] == 1 if jth regular class is a member of ith powerset class
            mapping[i, j] == 0 otherwise

        Example
        -------
        With num_classes == 3 and max_set_size == 2, returns

            [0, 0, 0]  # none
            [1, 0, 0]  # class #1
            [0, 1, 0]  # class #2
            [0, 0, 1]  # class #3
            [1, 1, 0]  # classes #1 and #2
            [1, 0, 1]  # classes #1 and #3
            [0, 1, 1]  # classes #2 and #3

        """
        mapping = np.zeros((self.num_powerset_classes, self.num_classes))
        
        powerset_k = 0
        for set_size in range(0, self.max_set_size + 1):
            for current_set in combinations(range(self.num_classes), set_size):
                mapping[powerset_k, current_set] = 1
                powerset_k += 1

        return mapping

    def build_cardinality(self) -> np.ndarray:
        """Compute size of each powerset class"""
        return np.sum(self.mapping, axis=1)

    def to_multilabel(self, powerset: np.ndarray, soft: bool = False) -> np.ndarray:
        """Convert predictions from powerset to multi-label

        Parameter
        ---------
        powerset : (batch_size, num_frames, num_powerset_classes) torch.Tensor
            Soft predictions in "powerset" space.
        soft : bool, optional
            Return soft multi-label predictions. Defaults to False (i.e. hard predictions)
            Assumes that `powerset` are "logits" (not "probabilities").

        Returns
        -------
        multi_label : (batch_size, num_frames, num_classes) torch.Tensor
            Predictions in "multi-label" space.
        """

        powerset_probs = np.identity(self.num_powerset_classes)[np.argmax(powerset, axis=-1)]
        return np.matmul(powerset_probs, self.mapping)


    def __call__(self, powerset: np.ndarray, soft: bool = False) -> np.ndarray:
        """Alias for `to_multilabel`"""
        
        return self.to_multilabel(powerset, soft=soft)