Skip to content

README.md

Latest commit

 

History

History
123 lines (101 loc) · 3.5 KB

README.md

File metadata and controls

123 lines (101 loc) · 3.5 KB

API Introduction

Overview

Setting Environment

If you want to use API, please add the repo_path/cpp and repo_path/python into environment path in the python file at first. Then import numpy and utils in repo_path/python.

import sys

dir_path = absolute_path_of_the_repo + "/cpp"

sys.path.append(dir_path)

dir_path = absolute_path_of_the_repo + "/python"
sys.path.append(dir_path)

import utils
import numpy

Functionality

Linear Conjugate Method

Given the quadratic function f(x):

$$\begin{aligned} f(x) = \frac{1}{2}x^{T}Ax + b^{T}x + c \end{aligned} $$

where $A$ is a positive-definite and symmetric matrix. Then linear CG method can find the global minima x_min by numerical method.

If you want to run the algorithm implemented by numpy, use the following function.

x_min = utils.np_linear_CG(x = initial_point, 
                        A = A, 
                        b = b, 
                        epsilon = convergence_tolerance, 
                        epoch = max_iteration)

If you want to run the algorithm implemented by c++, use the following function.
When num_threads > 1, some of the matrix operator would be paralleled, and the tiled matrix multiplication would be used.

x_min = utils.custom_linear_CG(x = initial_point, 
                            A = A, 
                            b = b, 
                            epsilon = convergence_tolerance, 
                            epoch = max_iteration, 
                            num_threads = number_of_threads)

Nonlinear Conjugate Method

Given the initial point x, convex function f(x) and it's derivative function df(x), the nonlinear CG method can find the local minima x_min by numerical method.

In order to get the df(x) for convenience, I suggest to use the autograd library.

from autograd import grad
import autograd.numpy as au

def example_function(x):
    """
    Functionality: f(x) = x1^4 + x2^4 + x3^4 + x4^4 ...+ xn^4
    Parameters: 
    x: The input vector.
    """
    x = np.array(x)
    return (np.sum(x**4))**0.5

example_derivative_function = utils.grad(example_function)

If you want to run the algorithm implemented by numpy, use the following function.

x_min = utils.np_nonlinear_CG(X = initial_point, 
    tol = convergence_tolerance, 
    alpha = hyperparameter_of_convergence_tolerance_in_line_search, 
    beta = convergence_rate_in_line_search, 
    f = function, 
    Df = derivative_function, 
    method = "Fletcher_Reeves")

If you want to run the algorithm implemented by c++, use the following function.
When num_threads > 1, some of the matrix operator would be paralleled, and the tiled matrix multiplication would be used.

x_min = utils.custom_nonlinear_CG(X = initial_point, 
    tol = convergence_tolerance, 
    alpha = hyperparameter_of_convergence_tolerance_in_line_search, 
    beta = convergence_rate_in_line_search, 
    f = function, 
    Df = derivative_function, 
    method = "Fletcher_Reeves",
    num_threads = number_of_threads)

By the way, there are three algorithm that the nonlinear CG method could choose, including "Fletcher_Reeves", "Dai-Yuan" and "Hager-Zhang". We can choose the algorithm by passing the string to the method parameter.