CNN MNIST handwriting recognition

#!/usr/bin/python3


from tensorflow.examples.tutorials.mnist import input_data  
import tensorflow as tf  
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)#  Read picture data set   
sess = tf.InteractiveSession()#  establish session  
#  One , Function declaration part   


def weight_variable(shape):  
    #  Normal distribution , The standard deviation is 0.1, Default maximum is 1, The minimum is -1, The mean for 0  
    initial = tf.truncated_normal(shape, stddev=0.1)  
    return tf.Variable(initial)  
def bias_variable(shape):  
    #  Create a structure for shape A matrix can also be said to be an array shape Declare its ranks , Initialize all values to 0.1  
    initial = tf.constant(0.1, shape=shape)  
    return tf.Variable(initial)  
def conv2d(x, W):    
    #  The number of convolution steps in each direction 1,SAME： Automatic filling outside the edge 0, Ergodic multiplication   
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')    
def max_pool_2x2(x):    
    #  Pooled convolution results （conv2d） The pool layer adopts kernel The size is 2*2, The number of steps is also 2, Peripheral supplement 0, Taking the maximum . The amount of data has shrunk 4 times   
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1], padding='SAME')    


#  Two , Define the input / output structure   


#  Declare a placeholder ,None Indicates that the number of input pictures is variable ,28*28 Image resolution   
xs = tf.placeholder(tf.float32, [None, 28*28])   
#  Category is 0-9 in total 10 Categories , Corresponding output classification result   
ys = tf.placeholder(tf.float32, [None, 10])   
keep_prob = tf.placeholder(tf.float32)  
# x_image Also put xs reshape a 28*28*1 The shape of the , Because it's a gray picture , So the passage is 1. As a training input,-1 The number of pictures is variable   
x_image = tf.reshape(xs, [-1, 28, 28, 1])   




#  3、 ... and , Build network , Define the algorithm formula , That is to say forward The calculation of time   


##  The first convolution operation  ##  
#  The first and second parameters are the convolution kernel size , namely patch, The third parameter is the number of image channels , The fourth parameter is the number of convolution kernels , Represents how many convolution feature images will appear ;  
W_conv1 = weight_variable([5, 5, 1, 32])   
#  For each convolution kernel, there is a corresponding offset .  
b_conv1 = bias_variable([32])    
#  Multiply the picture by the convolution kernel , And add paranoia , Convolution results 28x28x32  
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)    
#  Pooling results 14x14x32  Multiply the convolution result by the pooled convolution kernel   
h_pool1 = max_pool_2x2(h_conv1)   


##  Second layer convolution operation  ##     
# 32 Channel convolution , Convolution 64 Features     
w_conv2 = weight_variable([5,5,32,64])   
# 64 Paranoid data   
b_conv2  = bias_variable([64])   
#  Be careful h_pool1 It is the pool result of the upper layer ,# Convolution results 14x14x64  
h_conv2 = tf.nn.relu(conv2d(h_pool1,w_conv2)+b_conv2)    
#  Pooling results 7x7x64  
h_pool2 = max_pool_2x2(h_conv2)    
#  Original image size 28*28, The first image is reduced to 14*14, share 32 Zhang , After the second round, the image is reduced to 7*7, share 64 Zhang     


##  Third layer full connection operation  ##  
#  Two dimensional tensor , The first parameter 7*7*64 Of patch, It can also be thought that there is only one line 7*7*64 Convolution of data , The second parameter represents the total number of convolutions 1024 individual   
W_fc1 = weight_variable([7*7*64, 1024])   
# 1024 Paranoid data   
b_fc1 = bias_variable([1024])   
#  Pool the convolution result of the second layer reshape Cheng has only one line 7*7*64 Data # [n_samples, 7, 7, 64] ->> [n_samples, 7*7*64]  
h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])   
#  Convolution operation , The result is 1*1*1024, A single row multiplied by a single column equals 1*1 matrix ,matmul Realize the most basic matrix multiplication , differ tf.nn.conv2d Traversal multiplication of , It is automatically regarded as the forward vector and the backward vector   
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)   


# dropout operation , Reduce overfitting , In fact, it is to reduce the weight of some inputs on the upper layer scale, Even set it to 0, Increase the weight of some inputs , Even set it to 2, Prevent the evaluation curve from shaking , I think it is necessary when there are few samples   
#  Use placeholders , from dropout Automatically determine scale, You can also customize it , such as 0.5, according to tensorflow It can be seen from the documents , The actual value used in the program is 1/0.5=2, That is, some inputs are multiplied by 2, At the same time, some inputs are multiplied by 0  
keep_prob = tf.placeholder(tf.float32)   
h_fc1_drop = tf.nn.dropout(h_fc1,keep_prob) # Perform... On the convolution result dropout operation   


##  Fourth layer output operation  ##  
#  Two dimensional tensor ,1*1024 matrix convolution , common 10 A convolution , Corresponding to what we started ys The length is 10  
W_fc2 = weight_variable([1024, 10])    
b_fc2 = bias_variable([10])    
#  The final classification , The result is 1*1*10 softmax and sigmoid It's all based on logistic Classification algorithm , One is multi classification, the other is two classification   
y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)   


#  Four , Definition loss( Minimum error probability ), Select optimization optimization loss,  
cross_entropy = -tf.reduce_sum(ys * tf.log(y_conv)) #  Define the cross entropy as loss function     
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy) #  Call the optimizer to optimize , In fact, it is through feeding data to win cross_entropy To minimize the     


#  5、 ... and , Start data training and evaluation   
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(ys,1))  
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))  
tf.global_variables_initializer().run()  
for i in range(20000):  
    batch = mnist.train.next_batch(50)  
    if i%100 == 0:  
        train_accuracy = accuracy.eval(feed_dict={x:batch[0], ys: batch[1], keep_prob: 1.0})  
        print("step %d, training accuracy %g"%(i, train_accuracy))  
    train_step.run(feed_dict={x: batch[0], ys: batch[1], keep_prob: 0.5})  
print("test accuracy %g"%accuracy.eval(feed_dict={x: mnist.test.images, ys: mnist.test.labels, keep_prob: 1.0})) 

modified from http://blog.csdn.net/xukaiwen_2016/article/details/70880694