DNN基础教程

一、原理

前向传播：
http://ufldl.stanford.edu/wiki/index.php/%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C
反向传播：
http://blog.csdn.net/han_xiaoyang/article/details/50321873

二、Python实现

使用mnist数据集，测试python编写dnn的识别准确率。

mnist数据集是28*28维度的黑白图片，如下所示

mnist数据展示

# -*- coding: utf-8 -*-  
import cPickle

def load_data():
    with open('./mnist.pkl') as fp:
        training_data, valid_data, test_data = cPickle.load(fp)
    return training_data, valid_data, test_data   # training_data, valid_data, test_data 均是二元 tuple
if __name__ == '__main__':
     training_data, valid_data, test_data = load_data()
     print "the {0} size: {1}".format("training_data",len(training_data[0]))  
     print "the {0} size: {1}".format("valid_data",len(valid_data[0]))  
     print "the {0} size: {1}".format("test_data",len(test_data[0]))  
     print "the {0} size: {1}".format("image",len(training_data[0][0]))  

输出为：

the training_data size: 50000
the valid_data size: 10000
the test_data size: 10000
the image size: 784

mnist 数据读取，转为train test valid 以及对应的label

#!/usr/bin/env python  
# -*- coding: utf-8 -*-  

from numpy import *  
import numpy as np  
import cPickle  
def load_data():  
    """载入解压后的数据，并读取"""  
    with open('mnist.pkl','rb') as f:  
        try:  
            train_data,validation_data,test_data = cPickle.load(f)  
            print " the file open sucessfully"  
            # print train_data[0].shape  #(50000,784)  
            # print train_data[1].shape   #(50000,)  
            return (train_data,validation_data,test_data)  
        except EOFError:  
            print 'the file open error'  
            return None  

def data_transform():  
    """将数据转化为计算格式"""  
    t_d,va_d,te_d = load_data()  
    # print t_d[0].shape  # (50000,784)  
    # print te_d[0].shape  # (10000,784)  
    # print va_d[0].shape  # (10000,784)  
    # n1 = [np.reshape(x,784,1) for x in t_d[0]] # 将5万个数据分别逐个取出化成（784,1），逐个排列  
    n = [np.reshape(x, (784, 1)) for x in t_d[0]]  # 将5万个数据分别逐个取出化成（784,1），逐个排列  
    # print 'n1',n1[0].shape  
    # print 'n',n[0].shape  
    m = [vectors(y) for y in t_d[1]] # 将5万标签（50000,1）化为（10,50000）  
    train_data = zip(n,m)  # 将数据与标签打包成元组形式  
    n = [np.reshape(x, (784, 1)) for x in va_d[0]]  # 将5万个数据分别逐个取出化成（784,1），排列  
    validation_data = zip(n,va_d[1])   # 没有将标签数据矢量化  
    n = [np.reshape(x, (784, 1)) for x in te_d[0]]  # 将5万个数据分别逐个取出化成（784,1），排列  
    test_data = zip(n, te_d[1])  # 没有将标签数据矢量化  
    # print train_data[0][0].shape  #(784,）  
    # print "len(train_data[0])",len(train_data[0]) #2  
    # print "len(train_data[100])",len(train_data[100]) #2  
    # print "len(train_data[0][0])", len(train_data[0][0]) #784  
    # print "train_data[0][0].shape", train_data[0][0].shape #（784,1）  
    # print "len(train_data)", len(train_data)  #50000  
    # print train_data[0][1].shape  #(10,1)  
    # print test_data[0][1] # 7  
    return (train_data,validation_data,test_data)  
def vectors(y):  
    """赋予标签"""  
    label = np.zeros((10,1))  
    label[y] = 1.0 #浮点计算  
    return label  

DNN网络构造

#!/usr/bin/env python  
# -*- coding: utf-8 -*-  


import numpy as np  
import random  
class Network(object):   #默认为基类用于继承：print isinstance(network,object)  
    def __init__(self,sizes):  
        self.num_layers = len(sizes)  
        self.sizes = sizes  
        # print 'num_layers', self.num_layers  
        self.weight = [np.random.randn(a1, a2) for (a1, a2) in zip(sizes[1:], sizes[:-1])] #产生一个个数组  
        self.bias = [np.random.randn(a3,1) for a3 in sizes[1:]]  
        # print self.weight[0].shape  #(20,10)  

    def SGD(self,train_data,min_batch_size,epoches,eta,test_data=False):  
        """ 1) 打乱样本，将训练数据划分成小批次 
            2）计算出反向传播梯度 
            3） 获得权重更新"""  
        if test_data: n_test = len(test_data)  
        n = len(train_data)   #50000  
        random.shuffle(train_data)  # 打乱  
        min_batches = [train_data[k:k+min_batch_size] for k in xrange(0,n,min_batch_size)] #提取批次数据  
        for k in xrange(0,epoches):   #利用更新后的权值继续更新  
            random.shuffle(train_data)  # 打乱  
            for min_batch in min_batches:  #逐个传入，效率很低  
                self.updata_parameter(min_batch,eta)  
            if test_data:  
                num = self.evaluate(test_data)  
                print "the {0}th epoches: {1}/{2}".format(k,num,len(test_data))  
            else:  
                print 'epoches {0} completed'.format(k)  

    def forward(self,x):  
        """获得各层激活值"""  
        for w,b in zip(self.weight,self.bias):  
            x = sigmoid(np.dot(w, x)+b)  
        return x  

    def updata_parameter(self,min_batch,eta):  
        """1) 反向传播计算每个样本梯度值 
           2） 累加每个批次样本的梯度值 
           3） 权值更新"""  
        ndeltab = [np.zeros(b.shape) for b in self.bias]  
        ndeltaw = [np.zeros(w.shape) for w in self.weight]  
        for x,y in min_batch:  
            deltab,deltaw = self.backprop(x,y)  
            ndeltab = [nb +db for nb,db in zip(ndeltab,deltab)]  
            ndeltaw = [nw + dw for nw,dw in zip(ndeltaw,deltaw)]  
        self.bias = [b - eta * ndb/len(min_batch) for ndb,b in zip(ndeltab,self.bias)]  
        self.weight = [w - eta * ndw/len(min_batch) for ndw,w in zip(ndeltaw,self.weight)]  


    def backprop(self,x,y):  
        """执行前向计算，再进行反向传播，返回deltaw,deltab"""  
        # [w for w in self.weight]  
        # print 'len',len(w)  
        # print "self.weight",self.weight[0].shape  
        # print w[0].shape  
        # print w[1].shape  
        # print w.shape  
        activation = x  
        activations = [x]  
        zs = []  
        # feedforward  
        for w, b in zip(self.weight, self.bias):  
            # print w.shape,activation.shape,b.shape  
            z = np.dot(w, activation) +b  
            zs.append(z)   #用于计算f(z)导数  
            activation = sigmoid(z)  
            # print 'activation',activation.shape  
            activations.append(activation)  # 每层的输出结果  
        delta = self.top_subtract(activations[-1],y) * dsigmoid(zs[-1]) #最后一层的delta,np.array乘,相同维度乘  
        deltaw = [np.zeros(w1.shape) for w1 in self.weight]  #每一次将获得的值作为列表形式赋给deltaw  
        deltab = [np.zeros(b1.shape) for b1 in self.bias]  
        # print 'deltab[0]',deltab[-1].shape  
        deltab[-1] = delta  
        deltaw[-1] = np.dot(delta,activations[-2].transpose())  
        for k in xrange(2,self.num_layers):  
            delta = np.dot(self.weight[-k+1].transpose(),delta) * dsigmoid(zs[-k])  
            deltab[-k] = delta  
            deltaw[-k] = np.dot(delta,activations[-k-1].transpose())  
        return (deltab,deltaw)  

    def evaluate(self,test_data):  
        """评估验证集和测试集的精度,标签直接一个数作为比较"""  
        z = [(np.argmax(self.forward(x)),y) for x,y in test_data]  
        zs = np.sum(int(a == b) for a,b in z)  
        # zk = sum(int(a == b) for a,b in z)  
        # print "zs/zk:",zs,zk  
        return zs  

    def top_subtract(self,x,y):  
        return (x - y)  

def sigmoid(x):  
    return 1.0/(1.0+np.exp(-x))  

def dsigmoid(x):  
    z = sigmoid(x)  
    return z*(1-z)  

mnist 识别任务测试

#!/usr/bin/env python  
# -*- coding: utf-8 -*-  


import net_load_data  
# net_load_data.load_data()  
train_data,validation_data,test_data = net_load_data.data_transform()  

import net_network as net  
net1 = net.Network([784,30,10])  
min_batch_size = 10  
eta = 3.0  
epoches = 30  
net1.SGD(train_data,min_batch_size,epoches,eta,test_data)  
print "complete"