-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbpnet.py
193 lines (185 loc) · 8.07 KB
/
bpnet.py
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
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
import numpy as np
#import tensorflow as tf
from sklearn.preprocessing import StandardScaler
from sklearn.preprocessing import MinMaxScaler
import scipy.io as sio
import matplotlib.pyplot as plt
# matlab文件名
matfn = u'E:/matlabproj/spectra_data.mat'
data = sio.loadmat(matfn)
plt.close('all')
xi = data['NIR']
yi = data['octane']
xi=np.array(xi)
yi=np.array(yi)
print(xi.shape,yi.shape)
class bpnet(object):
def __init__(self,inputlayer,outputlayer,learnrate,hidelayer):#从一开始
self.inputlayer=inputlayer
self.outputlayer=outputlayer
self.learnrate=learnrate
self.hidelayer=hidelayer
self.ekout = np.zeros((self.outputlayer, self.inputlayer))
def addlayer(self):
self.weight=np.random.rand(self.hidelayer,self.inputlayer,self.inputlayer)#, dtype = float, order = 'C')#all weight,and hight=input highth每行是一个神经元
print("weight",self.weight)
self.bias=np.random.rand(self.hidelayer+1)#bias
self.outweight=np.random.rand(self.outputlayer,self.inputlayer)#, dtype = float, order = 'C')
self.__X=np.zeros((self.hidelayer,self.inputlayer),dtype = float, order = 'C')#中间的x=weight*input+bias
self.active=np.zeros((self.hidelayer,self.inputlayer),dtype = float, order = 'C')#激活值a
def sigmoid(self,x):
s = 1 / (1 + np.exp(-x))
return s
def sigmoidprime(self,x):
return np.multiply(x,(1-x))
def relu(self,x):
s = np.where(x < 0, 0, x)
return s
def ReLuPrime(self,x):
# ReLu 导数
#x[x <= 0] = 0
x[x > 0] = 1
return x
def forward(self,inputnum):
for i in range(self.hidelayer):#每一层的激活值存在active[i]里
#print('input',inputnum,self.weight[i],self.bias[i])
self.__X[i]=np.matmul(self.weight[i],inputnum)+self.bias[i]#
#print('X', i + 1, '层', self.__X[i])
self.active[i]=self.relu(self.__X[i])#self.sigmoid(self.__X[i])#1 / (1 + np.exp(-self.__X[i]))#激活值
#print('active',i+1,'层',self.active[i])
inputnum=self.active[i]#输入值得传递
outputnum=np.matmul(self.outweight,self.active[self.hidelayer-1])+self.bias[self.hidelayer]
return outputnum
def backforward(self,inputnum,outputParameter,outputnum):
#print("11212121212121111",self.ekout,outputParameter-outputnum,(outputParameter-outputnum)*self.outweight)
#self.ekout=np.zeros((self.outputlayer,self.inputlayer))
'''
for k in range(self.outputlayer):
self.ek[k] = outputParameter[k] - outputnum[k]'''
'''算误差'''
self.ekout=(outputParameter-outputnum)*self.outweight
#print("output err",self.ekout)
self.ekhide=np.zeros((self.hidelayer,self.inputlayer))
lastek=self.ekout[:]
for i in range(self.hidelayer):
self.ekhide[self.hidelayer-1-i]=np.matmul(self.weight[self.hidelayer-1-i],np.ones((self.inputlayer)))*lastek
'''print("误差值计算=",self.weight[self.hidelayer-1-i],"x",np.ones((self.inputlayer)),"=",np.matmul(self.weight[self.hidelayer-1-i],np.ones((self.inputlayer))))
print(self.ekhide[self.hidelayer - 1 - i], "=",
np.matmul(self.weight[self.hidelayer - 1 - i], np.ones((self.inputlayer))), "*", lastek)'''
lastek=self.ekhide[self.hidelayer-1-i][:]
#print("ekhide:",self.ekhide)
lastek=np.vstack((self.ekhide,self.ekout))
#print("eek:",lastek)
#self.J=np.sum(np.dot(self.ek,self.ek)/2.0)
'''算隐含层梯度'''
'''这个喵？'''
miao=np.vstack((inputnum,self.active[:self.hidelayer-1]))#,)
dweight0=self.learnrate*self.ekhide*self.active*self.ReLuPrime(self.active)#self.sigmoidprime(self.active)#(1-self.active)#乘输入值
self.dweight=np.ones((self.hidelayer,self.inputlayer,self.inputlayer))
#print('miao:',miao,self.dweight[0].shape)
#print('active:',self.active[:self.hidelayer-1])
#print('input',inputnum)
for k in range(self.hidelayer):
kk=dweight0[k].reshape(self.inputlayer,1)
self.dweight[k]=self.dweight[k]*miao[k]*kk
self.dweightout=self.learnrate*self.ekout*self.active[self.hidelayer-1]*outputnum*self.ReLuPrime(outputnum)#(1-outputnum)
#print("dweight",self.dweight)
#print("dweightout",self.dweightout)
'''for p in range(self.outputlayer):
self.J = self.J + self.ek[k] * self.ek[k] / 2.0#误差能量
for(int i=0;i<IM;i++){
for(int j=0;j<RM;j++)
{
for(int k=0;k<OM;k++)
{
dWin[i][j]=dWin[i][j]+learnRate*(Ek[k]*Wout[j][k]*XjActive[j]*(1-XjActive[j])*Xi[i]);
}//每个神经元的修正值
Win[i][j]=Win[i][j]+dWin[i][j]+alfa*(oldWin[i][j]-old1Win[i][j]);
old1Win[i][j]=oldWin[i][j];
oldWin[i][j]=Win[i][j];
}
}
self.dweight1=self.ek*inputnum*outputnum*self.active*(1-self.active)#每行是一层,神经元nxn的
dy=inputnum[:]
#self.dweight2 =np.zeros((self.hidelayer,self.inputlayer,self.inputlayer))
#for i in range(self.hidelayer):
# self.dweight2[i]=np.multiply(dy,self.dweight1[i])*self.learnrate
# dy=self.active[i]
#print("123",self.dweight2,self.dweight1)
#dy=self.active[self.hidelayer-1]
self.dweightout=self.ek*inputnum*(1-outputnum)*self.learnrate'''
'''算bias误差'''
#print("sum axis=1",np.sum(self.dweight,axis=(1,2)))
self.dbias=np.sum(self.dweight,axis=(1,2))*self.learnrate#每层神经元和
dd=np.sum(self.dweightout)
self.dbias=np.hstack((self.dbias,dd))
'''修正'''
self.weight-=self.dweight
#print("bias",self.bias,"dbias",self.dbias)
self.bias-=self.dbias
#print("bias",self.bias)
self.outweight-=self.dweightout
def train(self,inputnum,outputnum,s):
self.addlayer()
#inputnum, outputnum=self.onetrain(inputnum,outputnum)
for i in range(len(inputnum)):
mix=100
while(abs(mix)>s):
outp=self.forward(inputnum[i])
self.backforward(inputnum[i],outputnum[i],outp)
mix=np.sum(outputnum[i]-outp)
print("miss=",mix,outp,outputnum[i])
print("train:"+str(i))
def test(self,inputnum):
#inputnum=self.onetest(inputnum)
outputnum=np.ones(len(inputnum))
for i in range(len(inputnum)):
outputnum[i]=self.forward(inputnum[i])#self.oneback(self.forward(inputnum[i]))
print("喵喵",outputnum[i])#,inputnum[i])
return outputnum
'''
def onetrain(self,inputnum,outputnum):
self.min=np.min(inputnum,axis=1)
self.max=np.max(inputnum,axis=1)
for i in range(len(inputnum)):
inputnum[i]=(inputnum[i]-self.min[i])/(self.max[i]-self.min[i])
self.outmin = np.min(outputnum)
self.outmax = np.max(outputnum)
for i in range(np.size(outputnum,0)):
outputnum[i] = (outputnum[i] - self.outmin) / (self.outmax - self.outmin)
print(inputnum,outputnum)
return inputnum,outputnum
def onetest(self,inputnum):
tmin = np.min(inputnum, axis=1)
tmax = np.max(inputnum, axis=1)
for i in range(len(inputnum)):
inputnum[i]=(inputnum[i]-tmin[i])/(tmax[i]-tmin[i])
return inputnum
def oneback(self,outputnum):
for i in range(len(outputnum)):
outputnum[i]*=(self.outmax - self.outmin)
outputnum[i]+=+ self.outmin
return outputnum'''
print(yi.shape,yi,np.size(yi,0),yi[5])
''''''
'''归一化'''
mm = MinMaxScaler()
xi = mm.fit_transform(xi)
mm2=MinMaxScaler()
yi2 = mm2.fit_transform(yi)
yi2=yi2.reshape(-1)
yi=yi.reshape(-1)
print(xi,yi2)
bp1=bpnet(401,1,0.1,2)
print("training...")
bp1.train(xi[:50],yi2[:50],0.1)
print("test...")
yii=bp1.test(xi[50:])
yii=yii.reshape(10,1)
print(yii)
yii=mm2.inverse_transform(yii)
#除以最大最小误差-min (x-min)/(max-min)
print(yii,yi[15:20])
plt.plot(np.arange(10),yi[50:],'ro',np.arange(10),yii,'bs')
plt.show()
plt.savefig('figure2.png')