神经网络 相关代码

# 数值微分

import numpy as np

import matplotlib.pylab as plt

from mpl_toolkits.mplot3d import Axes3D

 

def _numerical_gradient_no_batch(f, x):

    h = 1e-4 # 0.0001

    grad = np.zeros_like(x)

 

    for idx in range(x.size):

        tem_val = x[idx]

        x[idx] = float(tem_val) + h

        fxh1 = f(x) # f(x+h)

 

        x[idx] = tem_val - h

        fxh2 = f(x) # f(x-h)

        grad[idx] = (fxh1 - fxh2) / (2 * h)

 

        x[idx] = tem_val # 还原值

 

    return grad

 

 

def numerical_gradient(f, X):

    if X.ndim == 1:

        return _numerical_gradient_no_batch(f, X)

    else:

        grad = np.zeros_like(X)

 

        for idx, x in enumerate(X):

            grad[idx] = _numerical_gradient_no_batch(f, x)

 

        return grad

 

 

def function_2(x):

    if x.ndim == 1:

        return np.sum(x ** 2)

    else:

        return np.sum(x ** 2, axis=1)

 

 

def tangent_line(f, x):

    d = numerical_gradient(f, x)

    print(d)

    y = f(x) - d * x

    return lambda t: d * t + y

 

 

if __name__ == '__main__':

    x0 = np.arange(-2, 2.5, 0.25)

    x1 = np.arange(-2, 2.5, 0.25)

    X, Y = np.meshgrid(x0, x1)

 

    X = X.flatten()

    Y = Y.flatten()

 

    grad = numerical_gradient(function_2, np.array([X, Y]))

 

    plt.figure()

    plt.quiver(X, Y, -grad[0], -grad[1], angles="xy",

               color="#666666") # ,headwidth = 10, scale = 40, color = "#444444")

    plt.xlim([-2, 2])

    plt.ylim([-2, 2])

    plt.xlabel('x0')

    plt.ylabel('x1')

    plt.grid()

    plt.legend()

    plt.draw()

    plt.show()

 

# 损失函数

# 均方误差

def mean_squared_error(y, t):

    return 0.5 * np.sum((y - t)**2)

 

# 交叉熵误差

# def cross_entropy_error(y, t):

# delta = 1e-7

# return -np.sum(t * np.log(y + delta))

 

# mini-batch 交叉熵误差

# y:神经网络输出，t:监督数据

def cross_entropy_