课程2

# -*- coding: utf-8 -*-
import numpy as np

def tanh(x):
    return np.tanh(x)

def tanh_deriv(x):
    return 1.0-np.tanh(x)*np.tanh(x)

def logistic(x):
    return 1/(1+np.exp(-x))

def logistic_derivate(x):
    return logistic(x)*(1-logistic(x))

class NeuralNetwork:
    def __init__(self, layers,activation='tanh'):
        #默认tanh
        #实例化一个对象，首先调用构造函数。self：指引类本身的指针，layer：包含神经网络有几层，每层有多少个神经元：10,2,2
        if activation == 'logistic':
            self.activation = logistic
            self.activation_deriv = logistic_derivative
        elif activation == 'tanh':
            self.activation = tanh
            self.activation_deriv = tanh_deriv

        self.weights=[]
        for i in range(1, len(layers)-1):
            #layer:10 10 3
            self.weights.append(2*np.random.random((layers[i-1]+1, layers[i]-1)*0.25)
            self.weights.append(2*np.random.random((layers[i]+1),layers[i+1]-1)*0.25)

    def fit(self, X, y, learning_rate=0.2, epochs=10000):#X 每一行对应一个实例，y 函数的分类标记
        #抽样的方法对神经网络进行更新，每抽取一次成为一个epochs循环
        #权重低于某个阈值；预测的错误率低于某个阈值；达到一定的循环次数。可以停止
        X = np.atleast_2d(X)
        temp = np.ones([X.shape[0],X.shape[1]+1]) #初始化矩阵都是1
        temp[:, 0:-1] #对bias偏置赋值,除了最后一列
        X= temp
        y = np.array(y)

        for k in range(epochs):
            i = np.random.randint(X.shape[0])#从0到n随机抽取一行
            a = [X[i]]
            for l in range(len(self.weights)):
                a.append(self.activation(np.dot(a[l], self.weights[l]))) #计算node 内积 完成正向所有更新
            error = y[i] - a[-1]  #真实y a最后一层 差值
            deltas = [error * self.activation_deriv(a[-1])]

            for l in range(len(a)-2,0,-1): #往回循环
                deltas.append(deltas[-1].dot(self.weights[l].T)*self.activation_derive)
            deltas.reverse()   #层数顺序颠倒
            for i in range(len(self.weights)):
                layer = np.atleast_2d(a[i])

                delta = np.atleast_2d(deltas[i])
                self.weights[i] += learning_rate *layer.T.dot(delta)

    def predit(self, x):
        x = np.array(x)
        temp = np.ones(x.shape[0]+1)
        temp[0:-1] = x
        a = temp
        for l in range(0, len(self.weights)):
            a = self.activation(np.dot(a,self.weights[l]))
        return a