1 '''
2 tensorflow 教程
3 mnist样例
4 '''
5 import tensorflow as tf
6 from tensorflow.examples.tutorials.mnist import input_data
7
8 #参数设置
9 INPUT_NODE=784
10 OUTPUT_NODE=10
11 LAYER1_NODE=500
12 BATCH_SIZE=100
13 LEARNING_RATE_BASE=0.8
14 LEARNING_RATE_DECAY=0.99
15 REGULARIZATION_RATE=0.0001
16 TRAINING_STEPS=10000
17 MOVEING_AVEARGE_DECAY=0.99
18
19
20 def inference(input_tensor,avg_class,weights1,biases1,weights2,biases2):
21 '''
22 定义前向计算的过程:
23 avg_class是滑动平均函数，使权重平滑过渡，保留历史数据，
24 为None时，表示普通的参数更新过程
25 '''
26 if avg_class==None:
27 layer1=tf.nn.relu(tf.matmul(input_tensor,weights1)+biases1)
28 return tf.matmul(layer1,weights2)+biases2
29 else:
30 layer1=tf.nn.relu(tf.matmul(input_tensor,avg_class.average(weights1)+avg_class.average(biases1)))
31 return tf.matmul(layer1,avg_class.average(weights2)+avg_class.average(biases2))
32
33 def train(mnist):
34 #设置输入变量 placerholder表示占位，开启会话训练的时候需要传入数据
35 x=tf.placeholder(tf.float32,[None,INPUT_NODE],name='x-input')
36 y_=tf.placeholder(tf.float32,[None,OUTPUT_NODE],name='y-input')
37
38 #设置权重变量，variable表示训练时需要自动更新
39 weights1=tf.Variable(tf.random_normal([INPUT_NODE,LAYER1_NODE],stddev=0.1))
40 biases1=tf.Variable(tf.constant(0.1,shape=[LAYER1_NODE]))
41 weights2=tf.Variable(tf.random_normal([LAYER1_NODE,OUTPUT_NODE],stddev=0.1))
42 biases2=tf.Variable(tf.constant(0.1,shape=[OUTPUT_NODE]))
43
44 #y=inference(x,None,weights1,biases1,weights2,biases2)
45
46 global_step=tf.Variable(0,trainable=False)#不可更新参数
47 variable_averages=tf.train.ExponentialMovingAverage(MOVEING_AVEARGE_DECAY,global_step)#min(decay,(1+step)/(10+step)) 后面的变量会越来越大，表示参数的更新越来越稳定，大都依赖于历史数据
48 variable_averages_op=variable_averages.apply(tf.trainable_variables())
49 average_y=inference(x,variable_averages,weights1,biases1,weights2,biases2)
50
51 cross_entropy=tf.nn.sparse_softmax_cross_entropy_with_logits(logits=average_y,labels=tf.argmax(y_,1))#计算图的输出是每个分类的得分，但是要求输入的标签是正确答案的下标
52 cross_entropy_mean=tf.reduce_mean(cross_entropy)
53
54 regularizer=tf.contrib.layers.l2_regularizer(REGULARIZATION_RATE)
55 regularization=regularizer(weights1)+regularizer(weights2)
56 loss=cross_entropy+regularization
57
58 learning_rate=tf.train.exponential_decay(LEARNING_RATE_BASE,global_step,mnist.train.num_examples/BATCH_SIZE,LEARNING_RATE_DECAY)#学习率成阶梯状衰减 每个epoch衰减一次，也就是一整轮数据训练完衰减一次
59 train_step=tf.train.GradientDescentOptimizer(learning_rate).minimize(loss,global_step=global_step)
60
61 train_op=tf.group(train_step,variable_averages_op)#把反向传播是需要更新的参数打包，不使用滑动平均不需要这句话，因为只更新权重。滑动平均还要利用历史数据更新并更新历史数据
62
63 correct_prediction=tf.equal(tf.argmax(average_y,1),tf.argmax(y_,1))
64 accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
65
66
67 with tf.Session() as sess:
68 tf.global_variables_initializer().run()
69 validate_feed={x:mnist.validation.images,y_:mnist.validation.labels}
70 test_feed={x:mnist.test.images,y_:mnist.test.labels}
71
72 for i in range(TRAINING_STEPS):
73 if i%1000==0:
74 validate_acc=sess.run(accuracy,feed_dict=validate_feed)
75 print('After %d training steps,validation accuracy using average model is %g' %(i,validate_acc))
76
77 xs,ys=mnist.train.next_batch(BATCH_SIZE)
78 sess.run(train_op,feed_dict={x:xs,y_:ys})
79
80 test_acc=sess.run(accuracy,feed_dict=test_feed)
81 print('After %d training steps,test accuracy using average model is %g' %(TRAINING_STEPS,test_acc))
82
83 def main(argv=None):
84 mnist=input_data.read_data_sets("/tmp/data",one_hot=True)
85 train(mnist)
86
87 if __name__ == '__main__':
88 tf.app.run()
