1 import tensorflow as tf
2 from tensorflow.examples.tutorials.mnist import input_data
3
4 INPUT_NODE = 784
5 OUTPUT_NODE = 10
6
7 # 1. 定义神经网络结构相关的参数
8 LAYER1_NODE = 500
9
10 BATCH_SIZE = 100
11
12 LEARNING_RATE_BASE = 0.8
13 LEARNING_RATE_DECAY = 0.99
14 REGULARIZATION_RATE = 0.0001
15 TRAINING_STEPS = 30000
16 MOVING_AVERAGE_DECAY = 0.99
17
18 def inference(input_tensor, avg_class, weights1, biases1, weights2, biases2):
19
20 if avg_class == None:
21
22 layer1 = tf.nn.relu(tf.matmul(input_tensor, weights1) + biases1)
23
24 return tf.matmul(layer1, weights2) + biases2
25
26 else:
27
28
29 layer1 = tf.nn.relu(tf.matmul(input_tensor, avg_class.average(weights1)) +
30 avg_class.average(biases1))
31 return tf.matmul(layer1, avg_class.average(weights2)) +
avg_class.average(biases2)
32
33 # 2. 定义训练过程
34 def train(mnist):
35 # 定义输入输出placeholder。
36 x = tf.placeholder(tf.float32, [None, INPUT_NODE], name=\'x-input\')
37 y_ = tf.placeholder(tf.float32, [None, OUTPUT_NODE], name=\'y-input\')
38
39 weights1 = tf.Variable(tf.truncated_normal([INPUT_NODE, LAYER1_NODE], stddev=0.1))
40 biases1 = tf.Variable(tf.constant(0.1, shape=[LAYER1_NODE]))
41
42 weights2 = tf.Variable(tf.truncated_normal([LAYER1_NODE, OUTPUT_NODE], stddev=0.1))
43 biases2 = tf.Variable(tf.constant(0.1, shape=[OUTPUT_NODE]))
44
45 y = inference(x, None, weights1, biases1, weights2, biases2)
46 global_step = tf.Variable(0, trainable=False)
47
48 # 定义损失函数、学习率、滑动平均操作以及训练过程。
49 variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY,
global_step)
50
51 variables_averages_op = variable_averages.apply(tf.trainable_variables())
52
53 average_y = inference(x, variable_averages, weights1, biases1, weights2, biases2)
54
55 cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits = y,
labels = tf.argmax(y_, 1))
56 cross_entropy_mean = tf.reduce_mean(cross_entropy)
57
58 regularizer = tf.contrib.layers.l2_regularizer(REGULARIZATION_RATE)
59 regularization = regularizer(weights1) + regularizer(weights2)
60
61 loss = cross_entropy_mean + regularization
62
63 learning_rate = tf.train.exponential_decay(
64 LEARNING_RATE_BASE,
65 global_step,
66 mnist.train.num_examples / BATCH_SIZE, LEARNING_RATE_DECAY)
67
68 train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss,
global_step=global_step)
69 with tf.control_dependencies([train_step, variables_averages_op]):
70 train_op = tf.no_op(name=\'train\')
71
72 correct_prediction = tf.equal(tf.argmax(average_y, 1), tf.argmax(y_,1))
73 accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
74
75 # 初始化TensorFlow持久化类。
76 with tf.Session() as sess:
77 tf.global_variables_initializer().run()
78 validate_feed = {x: mnist.validation.images,
79 y_: mnist.validation.labels}
80 test_feed = {x: mnist.test.images, y_: mnist.test.labels}
81
82 for i in range(TRAINING_STEPS):
83 if i % 1000 == 0:
84 validate_acc = sess.run(accuracy, feed_dict = validate_feed)
85 print("After %d training step(s), validation accuracy"
86 "using average model is %g " % (i, validate_acc))
87
88 xs, ys = mnist.train.next_batch(BATCH_SIZE)
89 sess.run(train_op, feed_dict={x: xs, y_: ys})
90
91 test_acc = sess.run(accuracy, feed_dict=test_feed)
92 print("After %d training step(s), test accuracy using average"
93 "model is %g" % (TRAINING_STEPS, test_acc))
94
95 # 3. 主程序入口
96 def main(argv=None):
97 mnist = input_data.read_data_sets("/tmp/data", one_hot=True)
98 train(mnist)
99
100 if __name__ == \'__main__\':
101 tf.app.run()
