機器學習中計算權重參數的幾類方法

# coding: utf-8
import numpy as np

def _numerical_gradient_1d(f, x):
    h = 1e-4 # 0.0001
    grad = np.zeros_like(x)
   
    for idx in range(x.size):
        tmp_val = x[idx]
        x[idx] = float(tmp_val) + h
        fxh1 = f(x) # f(x+h)
       
        x[idx] = tmp_val - h
        fxh2 = f(x) # f(x-h)
        grad[idx] = (fxh1 - fxh2) / (2*h)
       
        x[idx] = tmp_val # 還原值
       
    return grad

def numerical_gradient_2d(f, X):
    if X.ndim == 1:
        return _numerical_gradient_1d(f, X)
    else:
        grad = np.zeros_like(X)
       
        for idx, x in enumerate(X):
            grad[idx] = _numerical_gradient_1d(f, x)
       
        return grad

def numerical_gradient(f, x):
    h = 1e-4 # 0.0001
    grad = np.zeros_like(x)
   
    it = np.nditer(x, flags=['multi_index'], op_flags=['readwrite'])
    while not it.finished:
        idx = it.multi_index
        tmp_val = x[idx]
        x[idx] = float(tmp_val) + h
        fxh1 = f(x) # f(x+h)
       
        x[idx] = tmp_val - h
        fxh2 = f(x) # f(x-h)
        grad[idx] = (fxh1 - fxh2) / (2*h)
       
        x[idx] = tmp_val # 還原值
        it.iternext()  
       
    return grad