from tensorflow.examples.tutorials.mnist import input_data  
import tensorflow as tf  
  
#截断正态分布在均值附近  
def weight_variable(shape):  
    initial=tf.truncated_normal(shape,stddev=0.1)  
    return tf.Variable(initial)  
def bias_variable(shape):  
    initial=tf.constant(0.1,shape=shape)  
    return tf.Variable(initial)  
  
#步长1x1，用0填充使得卷积以后的向量维度和输入一样  
def conv2d(x,W):                                        
    return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding="SAME")  
  
#pooling核大小为2x2，步长为2x2pooling后向量维度减半  
def max_pool_2x2(x):                                     
    return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding="SAME")  
  
#data  
mnist = input_data.read_data_sets('./MNIST_data', one_hot=True)  
  
#placeholder  
x=tf.placeholder(tf.float32,shape=[None,784])  
y_=tf.placeholder(tf.float32,shape=[None,10])  
  
x_image=tf.reshape(x,[-1,28,28,1]) #将一维向量变为二维向量  
  
#conv layer  
W_conv1=weight_variable([5,5,1,32]) #第一个卷积层  
b_conv1=bias_variable([32])  
h_conv1=tf.nn.relu(conv2d(x_image,W_conv1)+b_conv1)  
h_pool1=max_pool_2x2(h_conv1)  
  
W_conv2=weight_variable([5,5,32,64]) #第二个卷积层  
b_conv2=bias_variable([64])  
h_conv2=tf.nn.relu(conv2d(h_pool1,W_conv2)+b_conv2)  
h_pool2=max_pool_2x2(h_conv2)  
  
#full connected layer  
W_fc1=weight_variable([7*7*64,1024])  #全连接层  
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)  
  
keep_prob=tf.placeholder("float")            #dropout  
h_fc1_drop=tf.nn.dropout(h_fc1,keep_prob)  
  
#model  output layer  
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)  
  
#loss  
cross_entropy=-tf.reduce_sum(y_*tf.log(y_conv))  
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)  
  
#evaluation  
#检测预测是否和真实标签匹配（索引位置一样表示匹配）  
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))   
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))  
  
# initialize and run cnn model  
saver=tf.train.Saver()  
  
with tf.Session() as sess:  
    sess.run(tf.global_variables_initializer())  
    for step in range(10):  
      for i in range(1100):  
          batch=mnist.train.next_batch(50)  
          if i%100==0:  
              train_accuracy=sess.run(accuracy,feed_dict={x:batch[0],y_:batch[1],keep_prob:1.0})  
              print("Iter: ", step, "step %d,training accuracy %g" % (i, train_accuracy))  
          train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})  
    print('test accurcay %g'%sess.run(accuracy,feed_dict=                          {x:mnist.test.images,y_:mnist.test.labels,keep_prob:1.0}))  
    saver.save(sess,r'model.ckpt')