利用ANN,BP实现简单的深度学习的代码:
#! /usr/bin/env python
# -*- coding=utf-8 -*-
import math
import random
import string
random.seed(0)
# 生成区间[a, b)内的随机数
def rand(a, b):
return (b - a) * random.random() + a
# 生成大小 I*J 的矩阵,默认零矩阵 (当然,亦可用 NumPy 提速)
def makeMatrix(I, J, fill=0.0):
m = []
for i in range(I):
m.append([fill] * J)
return m
# 函数 sigmoid,这里采用 tanh,因为看起来要比标准的 1/(1+e^-x) 漂亮些
def sigmoid(x):
return math.tanh(x)
# 函数 sigmoid 的派生函数, 为了得到输出 (即:y)
def dsigmoid(y):
return 1.0 - y ** 2
class NN:
''' 三层反向传播神经网络 '''
def __init__(self, ni, nh, no):
# 输入层、隐藏层、输出层的节点(数)
self.ni = ni + 1 # 增加一个偏差节点 #3
self.nh = nh #4
self.no = no #1
# 激活神经网络的所有节点(向量)
self.ai = [1.0] * self.ni #self.ni = 3,输入层有3个节点,初始化为[1.0, 1.0, 1.0]
self.ah = [1.0] * self.nh #self.nh = 4,隐藏层有4个节点,初始化位[1.0, 1.0, 1.0, 1.0]
self.ao = [1.0] * self.no #self.nh = 1,输出层有1个节点,初始化位[1.0]
# 建立权重(矩阵)
self.wi = makeMatrix(self.ni, self.nh)#输入层到隐藏层,3个到4个,权值:[[0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0]]
self.wo = makeMatrix(self.nh, self.no)#隐藏层到输出层,4个到1个,权值:[[0.0], [0.0], [0.0], [0.0]]
# 设为随机值
for i in range(self.ni):
for j in range(self.nh):
self.wi[i][j] = rand(-0.2, 0.2)
for j in range(self.nh):
for k in range(self.no):
self.wo[j][k] = rand(-2.0, 2.0)
print "随机权值"
print self.wi # 随机后的权值:[[0.13776874061001926, 0.10318176117612099, -0.031771367667662004, -0.09643329988281467], [0.004509888547443414, -0.03802634501983429, 0.11351943561390904, -0.07867490956842903], [-0.009361218339057675, 0.03335281578201249, 0.16324515407813406, 0.0018747423269561136]]
print self.wo #随机后的权值:[[-0.8726486224011847], [1.0232168166288957], [0.4734759867013265], [-0.9979746345502378]]
# 最后建立动量因子(矩阵)
self.ci = makeMatrix(self.ni, self.nh) #输入层到隐藏层,3个到4个,权值:[[0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0]]
self.co = makeMatrix(self.nh, self.no) #隐藏层到输出层,4个到1个,权值:[[0.0], [0.0], [0.0], [0.0]]
#更新的函数
def update(self, inputs):
print "-----------update---------------"
if len(inputs) != self.ni - 1:
raise ValueError('与输入层节点数不符!')
# 激活输入层
print "----------self.ai--------"
for i in range(self.ni - 1):
# self.ai[i] = sigmoid(inputs[i])
self.ai[i] = inputs[i]
print "jihuo_input"
print self.ai[i]
print "---------self.nh---------"
# 激活隐藏层
for j in range(self.nh):#self.nh = 4
sum = 0.0
for i in range(self.ni):#self.ni = 3
sum = sum + self.ai[i] * self.wi[i][j]
self.ah[j] = sigmoid(sum)
print self.ah[j]
# 激活输出层
print "---------self.ao---------"
for k in range(self.no): # self.no = 1
sum = 0.0
for j in range(self.nh): #self.nh = 4
sum = sum + self.ah[j] * self.wo[j][k]
self.ao[k] = sigmoid(sum)
print "---------"
print self.ao[:]
print "---------"
return self.ao[:]
#反向传播
def backPropagate(self, targets, N, M):
''' 反向传播 '''
print targets
if len(targets) != self.no:
raise ValueError('与输出层节点数不符!')
# 计算输出层的误差
output_deltas = [0.0] * self.no # 首先初始化误差值
for k in range(self.no): # self.no = 1
error = targets[k] - self.ao[k]
output_deltas[k] = dsigmoid(self.ao[k]) * error
# 计算隐藏层的误差 #首先初始化误差值
hidden_deltas = [0.0] * self.nh
for j in range(self.nh): #self.nh = 4
error = 0.0
for k in range(self.no): #self.no =1
error = error + output_deltas[k] * self.wo[j][k]
hidden_deltas[j] = dsigmoid(self.ah[j]) * error
# 更新输出层权重
for j in range(self.nh):
for k in range(self.no):
change = output_deltas[k] * self.ah[j]
self.wo[j][k] = self.wo[j][k] + N * change + M * self.co[j][k]
self.co[j][k] = change
# print(N*change, M*self.co[j][k])
# 更新输入层权重
for i in range(self.ni):
for j in range(self.nh):
change = hidden_deltas[j] * self.ai[i]
self.wi[i][j] = self.wi[i][j] + N * change + M * self.ci[i][j]
self.ci[i][j] = change
# 计算误差
error = 0.0
for k in range(len(targets)):
error = error + 0.5 * (targets[k] - self.ao[k]) ** 2
return error
def test(self, patterns):
for p in patterns:
print(p[0], '->', self.update(p[0]))
def weights(self):
print('输入层权重:')
for i in range(self.ni):
print(self.wi[i])
print()
print('输出层权重:')
for j in range(self.nh):
print(self.wo[j])
def train(self, patterns, iterations=1000, N=0.5, M=0.1):
# N: 学习速率(learning rate)
# M: 动量因子(momentum factor)
for i in range(iterations): #循环1000次,每100 次输出一个结果
print "---------------"
error = 0.0
for p in patterns:# 每次输入的有四个点
# print p
inputs = p[0]
# print "imputs"
# print inputs
# print "-------"
targets = p[1]
self.update(inputs) #更新输入的数据
error = error + self.backPropagate(targets, N, M) #反向传播结果,把结果回带到
print "******************************************"
if i % 100 == 0:
print('误差 %-.5f' % error)
def demo():
# 一个演示:教神经网络学习逻辑异或(XOR)------------可以换成你自己的数据试试
pat = [
[[0, 0], [0]],
[[0, 1], [1]],
[[1, 0], [1]],
[[1, 1], [0]]
]
# 创建一个神经网络:输入层有两个节点、隐藏层有两个节点、输出层有一个节点
n = NN(2, 4, 1)
# 用一些模式训练它
n.train(pat)
# 测试训练的成果(不要吃惊哦)
n.test(pat)
# 看看训练好的权重(当然可以考虑把训练好的权重持久化)
n.weights()
if __name__ == '__main__':
demo()