1 from tensorflow.examples.tutorials.mnist import input_data
 2 import tensorflow as tf
 3 
 4 mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
 5 sess = tf.InteractiveSession()
 6 
 7 def weight_variable(shape):
 8     '''初始化权重函数,truncated_normal创建标准差为0.2的截断正态函数'''
 9     initial = tf.truncated_normal(shape, stddev=0.1)
10     return tf.Variable(initial)
11 
12 def bias_variable(shape):
13     '''初始化偏置函数,由于使用ReLU要加一些正值0.1,避免死亡节点(dead neurons)'''
14     initial = tf.constant(0.1, shape = shape)
15     return tf.Variable(initial)
16 
17 def conv2d(x, W):
18     '''x:输入 w:卷积参数 例[5, 5, 1, 32]:5, 5为卷积核尺寸
19     1:为多少channel 彩色是3 灰度是1
20     32:为卷积核的数量(这个卷积层会提取的多少类特征)
21     strides:卷积模板移动的步长
22     padding:代表边界处理方式,SAME即输入和输出保持同样尺寸'''
23     return tf.nn.conv2d(x, W, strides = [1, 1, 1, 1], padding = 'SAME')
24 
25 def max_pool_2x2(x):
26     '''池化层函数 max_pool:最大池化函数'''
27     return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
28                           padding='SAME')
29 
30 #x:特征
31 x = tf.placeholder(tf.float32, [None, 784])
32 #y_真实的label
33 y_ = tf.placeholder(tf.float32, [None, 10])
34 '''卷积神经网络会利用到原有的空间结构信息,因此需要将1D的输入向量转化为2D图片结构(1x784->28x28)
35 因为只有一个颜色通道,故最终尺寸为[-1, 28, 28, 1]其中:-1代表样本数量不固定,1代表颜色通道数量'''
36 x_image = tf.reshape(x, [-1, 28, 28, 1])
37 
38 '''先定义weights和bias,然后使用conv2d函数进行卷积操作并加上偏置,
39 接着使用ReLU激活函数进行非线性处理,最好使用max_pool_2x2对卷积的输出结果进行池化操作'''
40 w_conv1 = weight_variable([5, 5, 1, 32])
41 b_conv1 = bias_variable([32])
42 h_conv1 = tf.nn.relu(conv2d(x_image, w_conv1)+b_conv1)
43 h_pool1 = max_pool_2x2(h_conv1)
44 
45 #定义第二个卷积层,不同在于特征变为64
46 w_conv2 = weight_variable([5, 5, 32, 64])
47 b_conv2 = bias_variable([64])
48 h_conv2 = tf.nn.relu(conv2d(h_pool1, w_conv2)+b_conv2)
49 h_pool2 = max_pool_2x2(h_conv2)
50 
51 '''经历两次2x2步长的最大池化,边长变为1/4,图片尺寸由28x28->7x7
52 由于第二个卷积层的卷积核数量为64,其输出tensor尺寸为7x7x64。
53 使用tf.reshape函数对其变形,转化为1D向量,然后连接一个全连接层,
54 隐含节点为1024,并使用Relu激活函数'''
55 w_fc1 = weight_variable([7*7*64, 1024])
56 b_fc1 = bias_variable([1024])
57 h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
58 h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, w_fc1)+b_fc1)
59 
60 #为减轻过拟合,使用一个dropout层
61 keep_prob = tf.placeholder(tf.float32)
62 h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
63 
64 #dropout层输出连softmax层,得到最后的概率输出
65 w_fc2 = weight_variable([1024, 10])
66 b_fc2 = bias_variable([10])
67 y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, w_fc2)+b_fc2)
68 
69 cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_*tf.log(y_conv),
70                                               reduction_indices=[1]))
71 
72 train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
73 
74 correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
75 accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
76 
77 tf.global_variables_initializer().run()
78 for i in range(20000):
79     batch = mnist.train.next_batch(50)
80     if i%100 == 0:
81         train_accuracy = accuracy.eval(feed_dict={x: batch[0], y_: batch[1], keep_prob: 1.0})
82         print("step %d, train accuracy %g"%(i, train_accuracy))
83 
84     train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
85 
86 print("test accuracy%g"%accuracy.eval(feed_dict={x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))
 1 step 0, train accuracy 0.22
 2 step 100, train accuracy 0.9
 3 step 200, train accuracy 0.88
 4 step 300, train accuracy 0.9
 5 step 400, train accuracy 0.96
 6 step 500, train accuracy 0.96
 7 step 600, train accuracy 0.98
 8 step 700, train accuracy 0.96
 9 step 800, train accuracy 0.98
10 step 900, train accuracy 0.96
11 step 1000, train accuracy 0.98
12 step 18000, train accuracy 1
13 step 18100, train accuracy 1
14 step 18200, train accuracy 0.98
15 step 18300, train accuracy 1
16 step 18400, train accuracy 1
17 step 18500, train accuracy 1
18 step 18600, train accuracy 1
19 step 18700, train accuracy 1
20 step 18800, train accuracy 1
21 step 18900, train accuracy 1
22 step 19000, train accuracy 1
23 step 19100, train accuracy 1
24 step 19200, train accuracy 1
25 step 19300, train accuracy 1
26 step 19400, train accuracy 1
27 step 19500, train accuracy 1
28 step 19600, train accuracy 1
29 step 19700, train accuracy 1
30 step 19800, train accuracy 1
31 step 19900, train accuracy 1
32 step 20000, train accuracy 1