1 #coding = utf-8
2 from tensorflow.examples.tutorials.mnist import input_data
3 import tensorflow as tf
4
5 mnist = input_data.read_data_sets("MNIST_data/", one_hot = True)
6 #创建一个IntercativeSession，这样后面的操作就无需指定Session
7 sess = tf.InteractiveSession()
8
9 #隐含层输出节点设置为300，（在此模型中隐含节点数设在200~1000结果区别不大）
10 in_units = 784
11 h1_units = 300
12 #利用tf.truncated_normal实现截断的正态分布，其标准差为0.1 [-1, 784]x[784, 300]
13 w1 = tf.Variable(tf.truncated_normal([in_units, h1_units], stddev=0.1))
14 b1 = tf.Variable(tf.zeros([h1_units]))
15 #[784, 300]x[300, 10]
16 w2 = tf.Variable(tf.zeros([h1_units, 10]))
17 b2 = tf.Variable(tf.zeros([10]))
18
19 #定义输入x，Dropout的比率keep_prob(通常在训练时小于1，而预测时等于1)
20 x = tf.placeholder(tf.float32, [None, in_units])
21 y_ = tf.placeholder(tf.float32, [None, 10])
22 keep_prob = tf.placeholder(tf.float32)
23
24 #hidden1:隐含层 y = relu(W1*x+b1)
25 hidden1 = tf.nn.relu(tf.matmul(x, w1) + b1)
26 '''调用tf.nn.dropout实现Dropout，keep_prob在训练时小于1，用于制造随机性，防止过拟合；
27 在预测时等于1，即使用全部特征来预测样本类别'''
28 hidden1_drop = tf.nn.dropout(hidden1, keep_prob)
29
30 #prediction
31 y = tf.nn.softmax(tf.matmul(hidden1_drop, w2)+b2)
32
33 cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_*tf.log(y), reduction_indices=[1]))
34 trian_step = tf.train.AdagradOptimizer(0.3).minimize(cross_entropy)
35
36 tf.global_variables_initializer().run()
37
38 for i in range(3000):
39 batch_xs, batch_ys = mnist.train.next_batch(100)
40 trian_step.run({x: batch_xs, y_: batch_ys, keep_prob: 0.75})
41 #out correct prediction
42 correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
43 accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
44 if i % 500 == 0:
45 print(accuracy.eval({x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))
1 0.2318
2 0.9584
3 0.9709
4 0.9761
5 0.9778
6 0.9782
