Source code for quara.minimization_algorithm.projected_gradient_descent_backtracking

import time
from typing import Callable, List

import numpy as np


from quara.loss_function.loss_function import LossFunction, LossFunctionOption
from quara.minimization_algorithm.projected_gradient_descent import (
    ProjectedGradientDescent,
    ProjectedGradientDescentOption,
    ProjectedGradientDescentResult,
)


[docs]class ProjectedGradientDescentBacktrackingResult(ProjectedGradientDescentResult): def __init__( self, value: np.ndarray, computation_time: float = None, k: int = None, fx: List[np.ndarray] = None, x: List[np.ndarray] = None, y: List[np.ndarray] = None, alpha: List[float] = None, error_values: List[float] = None, ): super().__init__(value, computation_time, k, fx, x, error_values) self._y: List[np.ndarray] = y self._alpha: List[float] = alpha @property def y(self) -> List[np.ndarray]: """return the y per iteration. Returns ------- List[np.ndarray] the y per iteration. """ return self._y @property def alpha(self) -> List[np.ndarray]: """return the alpha per iteration. Returns ------- List[np.ndarray] the alpha per iteration. """ return self._alpha
[docs]class ProjectedGradientDescentBacktrackingOption(ProjectedGradientDescentOption): def __init__( self, on_algo_eq_constraint: bool = True, on_algo_ineq_constraint: bool = True, var_start: np.ndarray = None, max_iteration_optimization: int = 1000, max_iteration_proj_physical: int = 100000, mu: float = None, gamma: float = 0.3, mode_stopping_criterion_gradient_descent: str = "single_difference_loss", num_history_stopping_criterion_gradient_descent: int = 1, mode_proj_order: str = "eq_ineq", eps: float = None, ): """Constructor Parameters ---------- on_algo_eq_constraint : bool, optional whether this algorithm needs on algorithm equality constraint, by default True on_algo_ineq_constraint : bool, optional whether this algorithm needs on algorithm inequality constraint, by default True var_start : np.ndarray, optional initial variable for the algorithm, by default None max_iteration_optimization: int, optional maximun number of iterations of optimization, by default 1000. max_iteration_proj_physical: int, optional maximun number of iterations of projection to physical, by default 100000. mu : float, optional algorithm option ``mu``, by default None gamma : float, optional algorithm option ``gamma``, by default 0.3 mode_stopping_criterion_gradient_descent : str, optional mode of stopping criterion for gradient descent, by default "single_difference_loss" num_history_stopping_criterion_gradient_descent : int, optional number of history to be used stopping criterion for gradient descent, by default 1 this must be a integer and greater than or equal to 1. mode_proj_order : str, optional the order in which the projections are performed, by default "eq_ineq". eps : float, optional algorithm option ``epsilon``, by default None """ super().__init__( on_algo_eq_constraint=on_algo_eq_constraint, on_algo_ineq_constraint=on_algo_ineq_constraint, var_start=var_start, max_iteration_optimization=max_iteration_optimization, max_iteration_proj_physical=max_iteration_proj_physical, mode_stopping_criterion_gradient_descent=mode_stopping_criterion_gradient_descent, num_history_stopping_criterion_gradient_descent=num_history_stopping_criterion_gradient_descent, mode_proj_order=mode_proj_order, eps=eps, ) if mu is None and var_start is not None: mu = 3 / (2 * np.sqrt(var_start.shape[0])) self._mu: float = mu self._gamma: float = gamma @property def mu(self) -> float: """returns algorithm option ``mu``. Returns ------- float algorithm option ``mu``. """ return self._mu @property def gamma(self) -> float: """returns algorithm option ``gamma``. Returns ------- float algorithm option ``gamma``. """ return self._gamma
[docs]class ProjectedGradientDescentBacktracking(ProjectedGradientDescent): def __init__(self, func_proj: Callable[[np.ndarray], np.ndarray] = None): """Constructor Parameters ---------- func_proj : Callable[[np.ndarray], np.ndarray], optional function of projection, by default None """ super().__init__(func_proj) self._is_gradient_required: bool = True self._is_hessian_required: bool = False
[docs] def is_loss_sufficient(self) -> bool: """returns whether the loss is sufficient. Returns ------- bool whether the loss is sufficient. """ if self.loss is None: return False elif self.loss.on_value is False: return False elif self.loss.on_gradient is False: return False else: return True
[docs] def is_option_sufficient(self) -> bool: """returns whether the option is sufficient. Returns ------- bool whether the option is sufficient. """ if self.option is None: return False elif self.option.mu is not None and self.option.mu <= 0: return False elif self.option.gamma is None or self.option.gamma <= 0: return False elif self.option.eps is None or self.option.eps <= 0: return False else: return True
[docs] def optimize( self, loss_function: LossFunction, loss_function_option: LossFunctionOption, algorithm_option: ProjectedGradientDescentBacktrackingOption, on_iteration_history: bool = False, ) -> ProjectedGradientDescentBacktrackingResult: """optimizes using specified parameters. Parameters ---------- loss_function : LossFunction Loss Function loss_function_option : LossFunctionOption Loss Function Option algorithm_option : ProjectedGradientDescentBacktrackingOption Projected Gradient Descent Backtracking Algorithm Option on_iteration_history : bool, optional whether to return iteration history, by default False Returns ------- ProjectedGradientDescentBacktrackingResult the result of the optimization. Raises ------ ValueError when ``on_value`` of ``loss_function`` is False. ValueError when ``on_gradient`` of ``loss_function`` is False. """ max_iteration = algorithm_option.max_iteration_optimization if loss_function.on_value == False: raise ValueError( "to execute ProjectedGradientDescentBase, 'on_value' of loss_function must be True." ) if loss_function.on_gradient == False: raise ValueError( "to execute ProjectedGradientDescentBase, 'on_gradient' of loss_function must be True." ) if algorithm_option.var_start is None: x_prev = ( self._qt.generate_empty_estimation_obj_with_setting_info() .generate_origin_obj() .to_var() ) else: x_prev = algorithm_option.var_start x_next = None if algorithm_option.mu: mu = algorithm_option.mu elif algorithm_option.var_start is not None: mu = 3 / (2 * np.sqrt(len(algorithm_option.var_start))) elif self._qt: mu = 3 / (2 * np.sqrt(self._qt.num_variables)) else: raise ValueError("unable to set the algorithm option mu.") gamma = algorithm_option.gamma eps = algorithm_option.eps # variables for debug if on_iteration_history: start_time = time.time() fxs = [loss_function.value(x_prev)] xs = [x_prev] ys = [] alphas = [] error_values = [] is_doing = True for k in range(1, max_iteration + 1): # shift variables if x_next is not None: x_prev = x_next y_prev = ( self.func_proj(x_prev - loss_function.gradient(x_prev) / mu) - x_prev ) alpha = 1.0 while self._is_doing_for_alpha(x_prev, y_prev, alpha, gamma, loss_function): alpha = 0.5 * alpha x_next = x_prev + alpha * y_prev # calc error value depend on "mode_stopping_criterion_gradient_descent" if ( algorithm_option.mode_stopping_criterion_gradient_descent == "single_difference_loss" ): error_value = loss_function.value(x_prev) - loss_function.value(x_next) elif ( algorithm_option.mode_stopping_criterion_gradient_descent == "sum_absolute_difference_loss" ): error_value = np.abs( loss_function.value(x_prev) - loss_function.value(x_next) ) elif ( algorithm_option.mode_stopping_criterion_gradient_descent == "sum_absolute_difference_variable" ): error_value = np.sqrt(np.sum((x_prev - x_next) ** 2)) elif ( algorithm_option.mode_stopping_criterion_gradient_descent == "sum_absolute_difference_projected_gradient" ): error_value = np.sqrt(np.sum(y_prev ** 2)) error_values.append(error_value) # calc sum of error values # if num_history_stopping_criterion_gradient_descent = 1, then this is single sum sum_range = min( len(error_values), algorithm_option.num_history_stopping_criterion_gradient_descent, ) value = np.sum(error_values[-sum_range:]) # variables for iteration history if on_iteration_history: fxs.append(loss_function.value(x_next, validate=True)) xs.append(x_next) ys.append(y_prev) alphas.append(alpha) is_doing = True if value > eps else False if not is_doing: break if k == max_iteration: start_red = "\033[31m" end_color = "\033[0m" print( f"{start_red}Warning!{end_color} pgdb iterations exceeds the limit {max_iteration}." ) if on_iteration_history: computation_time = time.time() - start_time result = ProjectedGradientDescentBacktrackingResult( x_next, computation_time=computation_time, k=k, fx=fxs, x=xs, y=ys, alpha=alphas, error_values=error_values, ) return result else: result = ProjectedGradientDescentBacktrackingResult(x_next) return result
def _is_doing_for_alpha( self, x_prev: np.ndarray, y_prev: np.ndarray, alpha: float, gamma: float, loss_function: LossFunction, ) -> bool: left_side = loss_function.value(x_prev + alpha * y_prev) right_side = loss_function.value(x_prev) + gamma * alpha * ( np.dot(y_prev, loss_function.gradient(x_prev)) ) return left_side > right_side