本文主要讲如何不依赖tenserflow等高级api实现一个简单的神经网络来做分类，所有的代码都在下面；在构造的数据（通过程序构造）上做了验证，经过1个小时的训练分类的准确率可以达到97%。
完整的结构化代码见于：链接地址
先来说说原理
网络构造
上面是一个简单的三层网络；输入层包含节点x1 , x2；隐层包含h1，h2；输出层包含o1。
输入节点的数量要等于输入数据的变量数目。
隐层节点的数量通过经验来确定。
如果只是做分类，输出层一般一个节点就够了。
从输入到输出的过程
1.输入节点的输出等于输入，x1节点输入x1时，输出还是x1.
2. 隐层和输出层的输入i为上层输出的加权求和再加偏置，输出为f(i) ， f为激活函数，可以取sigmoid。h1的输出为 sigmoid(w1x1 + w2x2 + b)
误差反向传播的过程
python实现
构造测试数据
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
|
# -*- coding: utf-8 -*-
import numpy as np
from random import random as rdn
'''
说明：我们构造1000条数据，每条数据有三个属性(用a1 , a2 , a3表示)
a1 离散型 取值 1 到 10 ， 均匀分布
a2 离散型 取值 1 到 10 ， 均匀分布
a3 连续型 取值 1 到 100 ， 且符合正态分布
各属性之间独立。
共2个分类(0 , 1)，属性值与类别之间的关系如下，
0 : a1 in [1 , 3] and a2 in [4 , 10] and a3 <= 50
1 : a1 in [1 , 3] and a2 in [4 , 10] and a3 > 50
0 : a1 in [1 , 3] and a2 in [1 , 3] and a3 > 30
1 : a1 in [1 , 3] and a2 in [1 , 3] and a3 <= 30
0 : a1 in [4 , 10] and a2 in [4 , 10] and a3 <= 50
1 : a1 in [4 , 10] and a2 in [4 , 10] and a3 > 50
0 : a1 in [4 , 10] and a2 in [1 , 3] and a3 > 30
1 : a1 in [4 , 10] and a2 in [1 , 3] and a3 <= 30
'''
def gendata() :
#为a3生成符合正态分布的数据
a3_data = np.random.randn(1000) * 30 + 50
data = []
for i in range(1000) :
#生成a1
a1 = int(rdn()*10) + 1
if a1 > 10 :
a1 = 10
#生成a2
a2 = int(rdn()*10) + 1
if a2 > 10 :
a2 = 10
#取a3
a3 = a3_data[i]
#计算这条数据对应的类别
c_id = 0
if a1 <= 3 and a2 >= 4 and a3 <= 50 :
c_id = 0
elif a1 <= 3 and a2 >= 4 and a3 > 50 :
c_id = 1
elif a1 <= 3 and a2 < 4 and a3 > 30 :
c_id = 0
elif a1 <= 3 and a2 < 4 and a3 <= 30 :
c_id = 1
elif a1 > 3 and a2 >= 4 and a3 <= 50 :
c_id = 0
elif a1 > 3 and a2 >= 4 and a3 > 50 :
c_id = 1
elif a1 > 3 and a2 < 4 and a3 > 30 :
c_id = 0
elif a1 > 3 and a2 < 4 and a3 <= 30 :
c_id = 1
else :
print('error')
#拼合成字串
str_line = str(i) + ',' + str(a1) + ',' + str(a2) + ',' + str(a3) + ',' + str(c_id)
data.append(str_line)
return '\n'.join(data)
|
激活函数
|
1
2
3
4
5
6
7
8
9
10
11
|
# -*- coding: utf-8 -*-
"""
created on sun dec 2 14:49:31 2018
@author: congpeiqing
"""
import numpy as np
#sigmoid函数的导数为 f(x)*(1-f(x))
def sigmoid(x) :
return 1/(1 + np.exp(-x))
|
网络实现
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
|
# -*- coding: utf-8 -*-
"""
created on sun dec 2 14:49:31 2018
@author: congpeiqing
"""
from activation_funcs import sigmoid
from random import random
class inputnode(object) :
def __init__(self , idx) :
self.idx = idx
self.output = none
def setinput(self , value) :
self.output = value
def getoutput(self) :
return self.output
def refreshparas(self , p1 , p2) :
pass
class neurode(object) :
def __init__(self , layer_name , idx , input_nodes , activation_func = none , powers = none , bias = none) :
self.idx = idx
self.layer_name = layer_name
self.input_nodes = input_nodes
if activation_func is not none :
self.activation_func = activation_func
else :
#默认取 sigmoid
self.activation_func = sigmoid
if powers is not none :
self.powers = powers
else :
self.powers = [random() for i in range(len(self.input_nodes))]
if bias is not none :
self.bias = bias
else :
self.bias = random()
self.output = none
def getoutput(self) :
self.output = self.activation_func(sum(map(lambda x : x[0].getoutput()*x[1] , zip(self.input_nodes, self.powers))) + self.bias)
return self.output
def refreshparas(self , err , learn_rate) :
err_add = self.output * (1 - self.output) * err
for i in range(len(self.input_nodes)) :
#调用子节点
self.input_nodes[i].refreshparas(self.powers[i] * err_add , learn_rate)
#调节参数
power_delta = learn_rate * err_add * self.input_nodes[i].output
self.powers[i] += power_delta
bias_delta = learn_rate * err_add
self.bias += bias_delta
class simplebp(object) :
def __init__(self , input_node_num , hidden_layer_node_num , trainning_data , test_data) :
self.input_node_num = input_node_num
self.input_nodes = [inputnode(i) for i in range(input_node_num)]
self.hidden_layer_nodes = [neurode('h' , i , self.input_nodes) for i in range(hidden_layer_node_num)]
self.output_node = neurode('o' , 0 , self.hidden_layer_nodes)
self.trainning_data = trainning_data
self.test_data = test_data
#逐条训练
def trainbyitem(self) :
cnt = 0
while true :
cnt += 1
learn_rate = 1.0/cnt
sum_diff = 0.0
#对于每一条训练数据进行一次训练过程
for item in self.trainning_data :
for i in range(self.input_node_num) :
self.input_nodes[i].setinput(item[i])
item_output = item[-1]
nn_output = self.output_node.getoutput()
#print('nn_output:' , nn_output)
diff = (item_output-nn_output)
sum_diff += abs(diff)
self.output_node.refreshparas(diff , learn_rate)
#print('refreshedparas')
#结束条件
print(round(sum_diff / len(self.trainning_data) , 4))
if sum_diff / len(self.trainning_data) < 0.1 :
break
def getaccuracy(self) :
cnt = 0
for item in self.test_data :
for i in range(self.input_node_num) :
self.input_nodes[i].setinput(item[i])
item_output = item[-1]
nn_output = self.output_node.getoutput()
if (nn_output > 0.5 and item_output > 0.5) or (nn_output < 0.5 and item_output < 0.5) :
cnt += 1
return cnt/(len(self.test_data) + 0.0)
|
主调流程
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
|
# -*- coding: utf-8 -*-
"""
created on sun dec 2 14:49:31 2018
@author: congpeiqing
"""
import os
from simplebp import simplebp
from gendata import gendata
if not os.path.exists('data'):
os.makedirs('data')
#构造训练和测试数据
data_file = open('data/trainning_data.dat' , 'w')
data_file.write(gendata())
data_file.close()
data_file = open('data/test_data.dat' , 'w')
data_file.write(gendata())
data_file.close()
#文件格式：rec_id,attr1_value,attr2_value,attr3_value,class_id
#读取和解析训练数据
trainning_data_file = open('data/trainning_data.dat')
trainning_data = []
for line in trainning_data_file :
line = line.strip()
fld_list = line.split(',')
trainning_data.append(tuple([float(field) for field in fld_list[1:]]))
trainning_data_file.close()
#读取和解析测试数据
test_data_file = open('data/test_data.dat')
test_data = []
for line in test_data_file :
line = line.strip()
fld_list = line.split(',')
test_data.append(tuple([float(field) for field in fld_list[1:]]))
test_data_file.close()
#构造一个二分类网络 输入节点3个，隐层节点10个，输出节点一个
simple_bp = simplebp(3 , 10 , trainning_data , test_data)
#训练网络
simple_bp.trainbyitem()
#测试分类准确率
print('accuracy : ' , simple_bp.getaccuracy())
#训练时长比较长，准确率可以达到97%
|
以上就是本文的全部内容，希望对大家的学习有所帮助，也希望大家多多支持服务器之家。
原文链接：https://blog.csdn.net/conggova/article/details/77799464