1 import tensorflow as tf
 2 from tensorflow.examples.tutorials.mnist import input_data
 3 
 4 # 下载训练和测试数据
 5 mnist = input_data.read_data_sets(‘MNIST_data/‘, one_hot = True)
 6 
 7 # 创建session
 8 sess = tf.Session()
 9 
10 # 占位符
11 x = tf.placeholder(tf.float32, shape=[None, 784]) # 每张图片28*28，共784个像素
12 y_ = tf.placeholder(tf.float32, shape=[None, 10]) # 输出为0-9共10个数字，其实就是把图片分为10类
13 
14 # 权重初始化
15 def weight_variable(shape):
16     initial = tf.truncated_normal(shape, stddev=0.1) # 使用截尾正态分布的随机数初始化权重，标准偏差是0.1（噪音）
17     return tf.Variable(initial)
18 
19 def bias_variable(shape):
20     initial = tf.constant(0.1, shape = shape) # 使用一个小正数初始化偏置，避免出现偏置总为0的情况
21     return tf.Variable(initial)
22 
23 # 卷积和集合
24 def conv2d(x, W): # 计算2d卷积
25     return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding=‘SAME‘)
26 
27 def max_pool_2x2(x): # 计算最大集合
28     return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding=‘SAME‘)
29 
30 # 第一层卷积
31 W_conv1 = weight_variable([5, 5, 1, 32]) # 为每个5*5小块计算32个特征
32 b_conv1 = bias_variable([32])
33 
34 x_image = tf.reshape(x, [-1, 28, 28, 1]) # 将图片像素转换为4维tensor，其中二三维是宽高，第四维是像素
35 h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
36 h_pool1 = max_pool_2x2(h_conv1)
37 
38 # 第二层卷积
39 W_conv2 = weight_variable([5, 5, 32, 64])
40 b_conv2 = bias_variable([64])
41 
42 h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
43 h_pool2 = max_pool_2x2(h_conv2)
44 
45 # 密集层
46 W_fc1 = weight_variable([7 * 7 * 64, 1024]) # 创建1024个神经元对整个图片进行处理
47 b_fc1 = bias_variable([1024])
48 
49 h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
50 h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
51 
52 # 退出（为了减少过度拟合，在读取层前面加退出层，仅训练时有效）
53 keep_prob = tf.placeholder(tf.float32)
54 h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
55 
56 # 读取层（最后我们加一个像softmax表达式那样的层）
57 W_fc2 = weight_variable([1024, 10])
58 b_fc2 = bias_variable([10])
59 
60 y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
61 
62 # 预测类和损失函数
63 cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv)) # 计算偏差平均值
64 train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy) # 每一步训练
65 
66 # 评估
67 correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
68 accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
69 sess.run(tf.global_variables_initializer())
70 
71 for i in range(1000):
72     batch = mnist.train.next_batch(50)
73     if i%10 == 0:
74         train_accuracy = accuracy.eval(feed_dict={ x:batch[0], y_: batch[1], keep_prob: 1.0}, session = sess) # 每10次训练计算一次精度
75         print("步数 %d, 精度 %g"%(i, train_accuracy))
76     train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5}, session = sess)
77 
78 # 关闭
79 sess.close()