过拟合：
真实的应用中，并不是让模型尽量模拟训练数据的行为，而是希望训练数据对未知做出判断。
模型过于复杂后，模型会积极每一个噪声的部分，而不是学习数据中的通用 趋势。当一个模型的参数比训练数据还要多的时候，这个模型就可以记忆这个所以训练数据的结果，而使损失函数为0.
避免过拟合的常用方法：正则化。在损失函数中加入刻画模型复杂程度的指标。损失函数：
J（θ）
引入正则化损失：J（θ）+λR(ω)
λ代表模型复杂损失在总损失的比列，R(ω)刻画的是模型的复杂程度。
模型的复杂程度由权重决定，一般。
常用的刻画模型复杂程度的函数R（ω）包括两种L1型正则和L2型正则。
loss=tf.ruduce_mean(tf.square(y_-y)+tf.contrib.layers.l2_ragularizer(lambda)(w))-----L2型正则的变量定义。、
除了引入正则化损失，还有避免过拟合的方法就是 加大训练数据 和 使用 Dropout。
1 import tensorflow as tf
2 from sklearn.datasets import load_digits #从sklearn的数据集引入手写字体数据集
3 from sklearn.model_selection import train_test_split # 作用：将数据集划分为 训练集和测试集
4 from sklearn.preprocessing import LabelBinarizer #数据的预处理
5 #binarizer二值化
6
7
8 ‘‘‘数据下载‘‘‘
9
10 digits=load_digits() #导入手写字体的datasets
11 X=digits.data #获得其特征向量
12 y=digits.target #获得样本label
13 y=LabelBinarizer().fit_transform(y) #二值化[0,1,1....]
14 ‘‘‘扩展sklearn.proprocessing.LabelBinarizer().fit_transform()‘‘‘
15 #lb=preprocessing.LabelBinarizer()
16 #>>> lb.fit_transform([‘yes‘, ‘no‘, ‘no‘, ‘yes‘])
17 #array([[1],
18 # [0],
19 # [0],
20 # [1]])
21 X_train,X_test,y_train,y_test =train_test_split(X,y,test_size=.3)
22 # test_size：float-获得多大比重的测试样本 （默认：0.25）---int - 获得多少个测试样本
23 #数据集被划分为训练集和测试集,label必须二值化因为分类的结果就是binarizer
24
25
26 ‘‘‘生成层 函数‘‘‘
27 def add_layer(input,in_size,out_size,n_layer=‘layer‘,activation_function=None):
28 layer_name=‘layer %s‘ % n_layer
29 ‘‘‘补充知识‘‘‘
30 #tf.name_scope:Wrapper for Graph.name_scope() using the default graph.
31 #scope名字的作用域
32 #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.
33 #If name has been used before, it will be made unique by calling self.unique_name(name).
34 with tf.name_scope(‘weights‘):
35 Weights=tf.Variable(tf.random_normal([in_size,out_size]),name=‘w‘)
36 tf.summary.histogram(layer_name+‘/wights‘,Weights)
37 #tf.summary.histogram:output summary with histogram直方图
38 #tf,random_normal正太分布
39 with tf.name_scope(‘biases‘):
40 biases=tf.Variable(tf.zeros([1,out_size])+0.1)
41 tf.summary.histogram(layer_name+‘/biases‘,biases)
42 #tf.summary.histogram:k
43 with tf.name_scope(‘Wx_plus_b‘):
44 Wx_plus_b=tf.matmul(input,Weights)+biases
45 ‘‘‘引入dropout,dropout添加在每一层的激活函数前‘‘‘
46 Wx_plus_b=tf.nn.dropout(Wx_plus_b,keep_prob)
47 #keep_prob 每个元素被留下来的概率
48 if activation_function==None:
49 outputs=Wx_plus_b
50 else:
51 outputs=activation_function(Wx_plus_b)
52 tf.summary.histogram(layer_name+‘/output‘,outputs)
53 return outputs
54 ‘‘‘准确率‘‘‘
55
56 ‘‘‘占位符‘‘‘
57 keep_prob = tf.placeholder(tf.float32)
58 xs=tf.placeholder(tf.float32,[None,64])
59 ys=tf.placeholder(tf.float32,[None,10])
60
61 ‘‘‘添加层‘‘‘
62 l1=add_layer(xs,64,50,‘l1‘,activation_function=tf.nn.softmax ) #据说其他激活函数会出错
63 prediction=add_layer(l1,50,10,activation_function=tf.nn.softmax)
64 #sotmax激活函数，用于分类函数
65
66 ‘‘‘计算‘‘‘
67 #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
68 ‘‘‘补充知识‘‘‘
69 #reduce_mean()
70 # ‘x‘ is [[1., 1. ]]
71 # [2., 2.]]
72 #tf.reduce_mean(x) ==> 1.5
73 #tf.reduce_mean(x, 0) ==> [1.5, 1.5]
74 #tf.reduce_mean(x, 1) ==> [1., 2.]
75 cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys*tf.log(prediction),reduction_indices=[1]))
76 tf.summary.scalar(‘loss‘,cross_entropy)
77 ‘‘‘补充知识‘‘‘
78 #reduce_sum
79 # ‘x‘ is [[1, 1, 1]]
80 # [1, 1, 1]]
81 #tf.reduce_sum(x) ==> 6
82 #tf.reduce_sum(x, 0) ==> [2, 2, 2]
83 #tf.reduce_sum(x, 1) ==> [3, 3]
84 #tf.reduce_sum(x, 1, keep_dims=True) ==> [[3], [3]]
85 #tf.reduce_sum(x, [0, 1]) ==> 6
86 train_step=tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)
87 merged=tf.summary.merge_all()
88 ‘‘‘Session_begin‘‘‘
89 with tf.Session() as sess:
90
91 #merged:tf.summary.merge_all:Merges all summaries collected in the default graph.
92 print("merged initialize sunccessfulliy")
93 train_writer=tf.summary.FileWriter(‘logs/train‘,sess.graph)
94 print("train_writer initialize sunccessfulliy")
95 test_writer=tf.summary.FileWriter(‘logs/test‘,sess.graph)
96 print("test_writer initialize sunccessfulliy")
97 sess.run(tf.global_variables_initializer())
98 print("variables initialize sunccessfulliy")
99 for i in range(1000):
100 #batch_xs,batch_ys=mnist.train.next_batch(100) #逐个batch去取数据
101 sess.run(train_step,feed_dict={xs:X_train,ys:y_train,keep_prob:0.6})
102 if(i%50==0):
103 #print(compute_accuracy(mnist.test.images,mnist.test.labels))
104 #train_result=sess.run(merged,feed_dict={xs:X_train,ys:y_train,keep_prob:1})
105 #test_result=sess.run(merged,feed_dict={xs:X_test,ys:y_test,keep_prob:1})
106 #train_writer.add_summary(train_result,i)
107 #test_writer.add_summary(test_result,i)
108 print(sess.run(cross_entropy,feed_dict={xs:X_test,ys:y_test,keep_prob:1}))
109 print("the {}".format(i))
110
原文地址：https://www.cnblogs.com/ChenKe-cheng/p/8893680.html
时间： 2024-10-09 13:17:33