DeepLearning&Tensorflow学习笔记2__mnist数据集CNN

1.Introduction

利用卷积神经网络CNN对Mnist数据集手写数字进行分类。

2.Source code

Version1:

coding:utf-8
import input_data
import tensorflow as tf
log_dir = "cnn_mnist_logs"
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
sess = tf.InteractiveSession()

x = tf.placeholder("float", shape=[None, 784])
y_ = tf.placeholder("float", shape=[None, 10])

#权重初始化
def weight_variable(shape):
  initial = tf.truncated_normal(shape, stddev=0.1)#使用截尾Guass分布初始化权重
  return tf.Variable(initial)
def bias_variable(shape):
  initial = tf.constant(0.1, shape=shape)
  return tf.Variable(initial)

#卷积和池化
def conv2d(x, W):
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
  #卷积使用1步长(stride size) 0边距(padding size的模板  保证输出和输入是同一个大小
  #par1:x表示input  #para1:batch_size #para2:in_height #para3:in_width  #para4:in_channels
  #par2:w表示filter #para1:filter_height #para2:filter_width #para3:in_channels  #para4:out_channels
  #par3:strides表示步长  #para1:batch_size #para2:height #para3:width  #para4:in_channels

def max_pool_2x2(x):   #池化用简单传统的2x2大小的模板做max pooling
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')
  #para1:x,      input
  #para2:ksize,  pooling size:2*2
  #para3:strdes, 步长2*2 [1,2,2,1]表示步长2×2 不重叠

#第一层卷积
W_conv1 = weight_variable([5, 5, 1, 32])  #前两个表示filter大小接着是输入的通道数目，最后是输出的通道数目;为每个5×5小块提取32个特征
b_conv1 = bias_variable([32])   #对于每一个输出通道都有一个对应的偏置量
x_image = tf.reshape(x, [-1,28,28,1]) #第2、第3维对应图片的宽、高，最后一维代表图片的颜色通道数(因为是灰度图所以这里的通道数为1，如果是rgb彩色图，则为3)

#We then convolve x_image with the weight tensor, add the bias, apply the ReLU function, and finally max pool
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)#利用ReLU激活函数   conv2d得到28×28×32
h_pool1 = max_pool_2x2(h_conv1)#池化  得到14×14×32

W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])

#第二层卷积
#为了构建一个更深的网络，我们会把几个类似的层堆叠起来。第二层中，每个5x5的patch会得到64个特征
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)   #得到14×14×64
h_pool2 = max_pool_2x2(h_conv2)  #得到7×7×64

#密集连接层
#现在，图片尺寸减小到7x7，我们加入一个有1024个神经元的全连接层，用于处理整个图片
#我们把池化层输出的张量reshape成一些向量，乘上权重矩阵，加上偏置，然后对其使用ReLU
W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])

h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

#Dropout
#减少过拟合，dropout可以屏蔽神经元的输出和某些突触
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

#输出层 最后，我们添加一个softmax层，就像前面的单层softmax regression一样
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)



#训练和评估模型
cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
train_writer = tf.summary.FileWriter(log_dir + '/train', sess.graph)   #定义写入摘要数据到事件日志的操作
sess.run(tf.initialize_all_variables())
for i in range(20000):   #20000
  batch = mnist.train.next_batch(50)
  if i%100 == 0:  #每100次训练计算一次训练集正确率
      train_accuracy = accuracy.eval(feed_dict={
         x:batch[0], y_: batch[1], keep_prob: 1.0})
      print "step %d, training accuracy %g"%(i, train_accuracy)
      print "test accuracy %g"%accuracy.eval(feed_dict={
           x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0})

  train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

#print "test accuracy %g"%accuracy.eval(feed_dict={
#    x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0})

Version2:

#coding:utf-8
import input_data
import tensorflow as tf
import numpy as np
log_dir = "cnn_mnist_logs"
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)


#权重初始化
def weight_variable(shape):
  initial = tf.truncated_normal(shape, stddev=0.1)#使用截尾Guass分布初始化权重
  return tf.Variable(initial)
def bias_variable(shape):
  initial = tf.constant(0.1, shape=shape)
  return tf.Variable(initial)

#卷积和池化
def conv2d(x, W):
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
  #卷积使用1步长(stride size) 0边距(padding size的模板  保证输出和输入是同一个大小
  #par1:x表示input  #para1:batch_size #para2:in_height #para3:in_width  #para4:in_channels
  #par2:w表示filter #para1:filter_height #para2:filter_width #para3:in_channels  #para4:out_channels
  #par3:strides表示步长  #para1:batch_size #para2:height #para3:width  #para4:in_channels

def max_pool_2x2(x):   #池化用简单传统的2x2大小的模板做max pooling
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')
  #para1:x,      input
  #para2:ksize,  pooling size:2*2       
  #para3:strdes, 步长2*2 [1,2,2,1]表示步长2×2 不重叠


graph = tf.Graph()
with graph.as_default():

        x = tf.placeholder("float", shape=[None, 784])
        y_ = tf.placeholder("float", shape=[None, 10])
        global_step = tf.Variable(0, trainable=False)

        #第一层卷积
        W_conv1 = weight_variable([5, 5, 1, 32])  #前两个表示filter大小接着是输入的通道数目，最后是输出的通道数目;为每个5×5小块提取32个特征
        b_conv1 = bias_variable([32])   #对于每一个输出通道都有一个对应的偏置量
        x_image = tf.reshape(x, [-1,28,28,1]) #第2、第3维对应图片的宽、高，最后一维代表图片的颜色通道数(因为是灰度图所以这里的通道数为1，如果是rgb彩色图，则为3)

        #We then convolve x_image with the weight tensor, add the bias, apply the ReLU function, and finally max pool
        h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)#利用ReLU激活函数   conv2d得到28×28×32
        h_pool1 = max_pool_2x2(h_conv1)#池化  得到14×14×32

        W_conv2 = weight_variable([5, 5, 32, 64])
        b_conv2 = bias_variable([64])

        #第二层卷积
        #为了构建一个更深的网络，我们会把几个类似的层堆叠起来。第二层中，每个5x5的patch会得到64个特征
        h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)   #得到14×14×64
        h_pool2 = max_pool_2x2(h_conv2)  #得到7×7×64

        #密集连接层
        #现在，图片尺寸减小到7x7，我们加入一个有1024个神经元的全连接层，用于处理整个图片
        #我们把池化层输出的张量reshape成一些向量，乘上权重矩阵，加上偏置，然后对其使用ReLU
        W_fc1 = weight_variable([7 * 7 * 64, 512])
        b_fc1 = bias_variable([512])

        h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
        h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

        #Dropout
        #减少过拟合，dropout可以屏蔽神经元的输出和某些突触
        keep_prob = tf.placeholder("float")
        #keep_prob=tf.Variable(tf.constant(0.5))
        h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

        #输出层 最后，我们添加一个softmax层，就像前面的单层softmax regression一样
        W_fc2 = weight_variable([512, 10])
        b_fc2 = bias_variable([10])
        y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)



        #训练和评估模型
        cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv))  #交叉熵
        optimizer = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
        predict_op = tf.argmax(y_conv, 1)  #放回py_x序列中最大值的index，概率最大的那>个，即类别（0--9）
        merged = tf.summary.merge_all()   ##定义合并变量操作，一次性生成所有摘要数据i
        correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))#accuracy分类正确率
        with tf.device("/cpu:0"):
           saver = tf.train.Saver(tf.global_variables(), max_to_keep=2)

with tf.Session(graph=graph) as sess:  #启动模型

  train_writer = tf.summary.FileWriter(log_dir + '/train', sess.graph)   #定义写入摘要数据到事件日志的操作
  tf.global_variables_initializer().run()   #将所有的变量初始化
  for i in range(20000):   #20000
    batch = mnist.train.next_batch(50)
    if i%100 == 0:  #每100次训练计算一次训练集正确率
      train_accuracy = accuracy.eval(feed_dict={
         x:batch[0], y_: batch[1], keep_prob: 1.0})
      print "step %d, training accuracy %g"%(i, train_accuracy)
      print "test accuracy %g"%accuracy.eval(feed_dict={
           x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0})

    optimizer.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
  saver.save(sess, 'run/checkpoint', global_step=global_step)   #使用tf.save.Saver保存整个模型后,让TensorBoard自动对模型中所有二维的Variable进行可视化

3.CNN Network framework