1 import tensorflow as tf
2 from tensorflow.examples.tutorials.mnist import input_data
3
4 '''数据下载'''
5 mnist=input_data.read_data_sets('Mnist_data',one_hot=True)
6 #one_hot标签
7
8 '''生成层 函数'''
9 def add_layer(input,in_size,out_size,n_layer='layer',activation_function=None):
10 layer_name='layer %s' % n_layer
11 '''补充知识'''
12 #tf.name_scope:Wrapper for Graph.name_scope() using the default graph.
13 #scope名字的作用域
14 #sprase:A string (not ending with '/') will create a new name scope, in which name is appended to the prefix of all operations created in the context.
15 #If name has been used before, it will be made unique by calling self.unique_name(name).
16 with tf.name_scope('weights'):
17 Weights=tf.Variable(tf.random_normal([in_size,out_size]),name='w')
18 tf.summary.histogram(layer_name+'/wights',Weights)
19 #tf.summary.histogram:output summary with histogram直方图
20 #tf,random_normal正太分布
21 with tf.name_scope('biases'):
22 biases=tf.Variable(tf.zeros([1,out_size])+0.1)
23 tf.summary.histogram(layer_name+'/biases',biases)
24 #tf.summary.histogram:k
25 with tf.name_scope('Wx_plus_b'):
26 Wx_plus_b=tf.matmul(input,Weights)+biases
27 if activation_function==None:
28 outputs=Wx_plus_b
29 else:
30 outputs=activation_function(Wx_plus_b)
31 tf.summary.histogram(layer_name+'/output',outputs)
32 return outputs
33 '''准确率'''
34 def compute_accuracy(v_xs,v_ys):
35 global prediction
36 y_pre=sess.run(prediction,feed_dict={xs:v_xs})#<
37 #tf.equal()对比预测值的索引和实际label的索引是否一样，一样返回True,否则返回false
38 correct_prediction=tf.equal(tf.argmax(y_pre,1),tf.argmax(v_ys,1))
39 #correct_prediction-->[ True False True ..., True True True]
40 '''补充知识-tf.argmax'''
41 #tf.argmax:Returns the index with the largest value across dimensions of a tensor.
42 #tf.argmax()----->
43 accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
44 #正确cast为1，错误cast为0
45 '''补充知识 tf.cast'''
46 #tf.cast: Casts a tensor to a new type.
47 ## tensor `a` is [1.8, 2.2], dtype=tf.float
48 #tf.cast(a, tf.int32) ==> [1, 2] # dtype=tf.int32
49 result=sess.run(accuracy,feed_dict={xs:v_xs,ys:v_ys})
50 #print(sess.run(correct_prediction,feed_dict={xs:v_xs,ys:v_ys}))
51 #ckc=tf.cast(correct_prediction,tf.float32)
52 #print(sess.run(ckc,feed_dict={xs:v_xs,ys:v_ys}))
53 return result
54
55
56 '''占位符'''
57 xs=tf.placeholder(tf.float32,[None,784])
58 ys=tf.placeholder(tf.float32,[None,10])
59
60 '''添加层'''
61
62 prediction=add_layer(xs,784,10,activation_function=tf.nn.softmax)
63 #sotmax激活函数，用于分类函数
64
65 '''计算'''
66 #交叉熵cross_entropy损失函数，参数分别为实际的预测值和实际的label值y，re
67 '''补充知识'''
68 #reduce_mean()
69 # 'x' is [[1., 1. ]]
70 # [2., 2.]]
71 #tf.reduce_mean(x) ==> 1.5
72 #tf.reduce_mean(x, 0) ==> [1.5, 1.5]
73 #tf.reduce_mean(x, 1) ==> [1., 2.]
74 cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys*tf.log(prediction),reduction_indices=[1]))
75 '''补充知识'''
76 #reduce_sum
77 # 'x' is [[1, 1, 1]]
78 # [1, 1, 1]]
79 #tf.reduce_sum(x) ==> 6
80 #tf.reduce_sum(x, 0) ==> [2, 2, 2]
81 #tf.reduce_sum(x, 1) ==> [3, 3]
82 #tf.reduce_sum(x, 1, keep_dims=True) ==> [[3], [3]]
83 #tf.reduce_sum(x, [0, 1]) ==> 6
84
85 train_step=tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)
86
87 '''Session_begin'''
88 with tf.Session() as sess:
89 sess.run(tf.global_variables_initializer())
90 for i in range(1000):
91 batch_xs,batch_ys=mnist.train.next_batch(100) #逐个batch去取数据
92 sess.run(train_step,feed_dict={xs:batch_xs,ys:batch_ys})
93 if(i%50==0):
94 print(compute_accuracy(mnist.test.images,mnist.test.labels))
95
