Module spatial_inequality.optimization.early_stopper

Implements an "early stopping" termination criterion for our algorithm, preventing unnecesary computations when improvements are no longer observed.

"""
Implements an "early stopping" termination criterion for our algorithm,
preventing unnecesary computations when improvements are no longer observed.
"""
import numpy as np

class EarlyStopper:
    """
    Class to handle preemptive stopping of an algorithm when no measurable
    improvement has been registered for a given amount of iterations.

    Attributes:
        __n_iterations (int): Number of allowed iterations without improvement.
        __tolerance (float): Tolerance for floating point measurement
            comparison.
        __absolute_min (float): Minimum value registered over all iterations.
        __n_iterations_wo_improvement (int): Current number of iterations with
            no observed improvement.
    """
    __n_iterations = None
    __tolerance = None
    
    __absolute_min = None
    __n_iterations_wo_improvement = None
    
    def __init__(self, n_iterations, tolerance=0.1):
        self.__n_iterations = n_iterations
        self.__tolerance = tolerance

        self.__absolute_min = None
        self.__n_iterations_wo_improvement = 0
        
    def update(self, value):
        """
        Records a new value for the tracked metric and updates all variables
        accordingly.

        Args:
            value (float): Updated value for tracked metric.

        Raises:
            StopIteration: Whenever more than the specified maximum amount of
                iterations has elapsed wihtout a measurable improvement to the
                tracked metric.
        """
        if self.__absolute_min is None:
            self.__absolute_min = value
        else:
            is_close = np.isclose(value, self.__absolute_min, atol=self.__tolerance)
            is_greater = value > self.__absolute_min
            if is_close or is_greater:
                # No detected improvement
                self.__n_iterations_wo_improvement += 1
            else:
                # Improvement detected
                self.__absolute_min = value
                self.__n_iterations_wo_improvement = 0
        if self.__n_iterations_wo_improvement >= self.__n_iterations:
            raise StopIteration

Classes

class EarlyStopper (n_iterations, tolerance=0.1)

Class to handle preemptive stopping of an algorithm when no measurable improvement has been registered for a given amount of iterations.

Attributes

__n_iterations : int

Number of allowed iterations without improvement.

__tolerance : float

Tolerance for floating point measurement comparison.

__absolute_min : float

Minimum value registered over all iterations.

__n_iterations_wo_improvement : int

Current number of iterations with no observed improvement.

Expand source code

class EarlyStopper:
    """
    Class to handle preemptive stopping of an algorithm when no measurable
    improvement has been registered for a given amount of iterations.

    Attributes:
        __n_iterations (int): Number of allowed iterations without improvement.
        __tolerance (float): Tolerance for floating point measurement
            comparison.
        __absolute_min (float): Minimum value registered over all iterations.
        __n_iterations_wo_improvement (int): Current number of iterations with
            no observed improvement.
    """
    __n_iterations = None
    __tolerance = None
    
    __absolute_min = None
    __n_iterations_wo_improvement = None
    
    def __init__(self, n_iterations, tolerance=0.1):
        self.__n_iterations = n_iterations
        self.__tolerance = tolerance

        self.__absolute_min = None
        self.__n_iterations_wo_improvement = 0
        
    def update(self, value):
        """
        Records a new value for the tracked metric and updates all variables
        accordingly.

        Args:
            value (float): Updated value for tracked metric.

        Raises:
            StopIteration: Whenever more than the specified maximum amount of
                iterations has elapsed wihtout a measurable improvement to the
                tracked metric.
        """
        if self.__absolute_min is None:
            self.__absolute_min = value
        else:
            is_close = np.isclose(value, self.__absolute_min, atol=self.__tolerance)
            is_greater = value > self.__absolute_min
            if is_close or is_greater:
                # No detected improvement
                self.__n_iterations_wo_improvement += 1
            else:
                # Improvement detected
                self.__absolute_min = value
                self.__n_iterations_wo_improvement = 0
        if self.__n_iterations_wo_improvement >= self.__n_iterations:
            raise StopIteration

Methods

def update(self, value)

Records a new value for the tracked metric and updates all variables accordingly.

Args

value : float

Updated value for tracked metric.

Raises

StopIteration

Whenever more than the specified maximum amount of iterations has elapsed wihtout a measurable improvement to the tracked metric.

Expand source code

def update(self, value):
    """
    Records a new value for the tracked metric and updates all variables
    accordingly.

    Args:
        value (float): Updated value for tracked metric.

    Raises:
        StopIteration: Whenever more than the specified maximum amount of
            iterations has elapsed wihtout a measurable improvement to the
            tracked metric.
    """
    if self.__absolute_min is None:
        self.__absolute_min = value
    else:
        is_close = np.isclose(value, self.__absolute_min, atol=self.__tolerance)
        is_greater = value > self.__absolute_min
        if is_close or is_greater:
            # No detected improvement
            self.__n_iterations_wo_improvement += 1
        else:
            # Improvement detected
            self.__absolute_min = value
            self.__n_iterations_wo_improvement = 0
    if self.__n_iterations_wo_improvement >= self.__n_iterations:
        raise StopIteration