/
MMNN.py
172 lines (150 loc) · 5.22 KB
/
MMNN.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
# -*- coding: utf-8 -*-
"""
Created on Sat Apr 27 23:46:54 2019
@author: jianfal
"""
from gen_data import gen_data
import torch as pt
import numpy as np
from sklearn.model_selection import train_test_split
class MLP(pt.nn.Module):
def __init__(self):
super(MLP,self).__init__()
self.fc1 = pt.nn.Linear(5,2)
self.fc2 = pt.nn.Linear(2,1)
def forward(self,din,random_effect):
dout = self.fc1(din)
dout = self.fc2(dout)
return dout+random_effect
def AccuarcyCompute(pred,label):
pred = pred.data.numpy()
label = label.data.numpy()
var=np.mean((pred-label)**2)
error=np.mean(abs(pred-label))
error_percentage=round(error/np.mean(abs(label))*100,2)
return var,error,error_percentage
def E_step(sigma_u,sigma_e,Y,Z):
K=Z.shape[1]
Sigma=(1/sigma_u)*(np.ones(K))
Sigma=Sigma+np.dot(Z.T,Z).diagonal()/sigma_e
Sigma1=1/(Sigma)
mu=np.zeros(K)
for i in range(len(Z)):
mu=mu+Sigma1*Z[i,]*Y[i,]/sigma_e
return np.reshape(mu,(K,1)),np.reshape(Sigma1,(K,1))
def M_step(mu,Sigma,Y,Z):
K=Z.shape[1]
sigma_u=(np.dot(mu.T,mu)+np.sum(Sigma))/K
sigma_e=np.sum((Y)**2)
for i in range(len(Z)):
sigma_e=sigma_e+np.dot(Z[i,],mu*mu+Sigma)-2*np.dot(Y[i,],np.dot(Z[i,],mu))
sigma_e=sigma_e/len(Z)
return sigma_u,sigma_e
def EMNN(y):
#pre_set
sigma_u=1
sigma_e=1
x_train, x_test, y_train, y_test, z_train, z_test = train_test_split(Xdata[:,:5], y,Xdata[:,5:], test_size = 0.3)
model = MLP()
optimizer = pt.optim.SGD(model.parameters(),lr=0.01,momentum=0.9)
lossfunc = pt.nn.MSELoss()
#training data
X=x_train
Y=y_train
Z=z_train
random_effect=np.zeros((len(X),1))
def NN(x,y,random_effect):
for i in range(100):
optimizer.zero_grad()
(inputs,labels) = x,y
inputs=pt.from_numpy(inputs)
labels=pt.from_numpy(labels)
inputs = pt.autograd.Variable(inputs)
labels = pt.autograd.Variable(labels)
random_e=pt.from_numpy(random_effect).float()
outputs = model(inputs.float(),random_e)
loss = lossfunc(outputs,labels.float())
loss.backward()
optimizer.step()
return model(inputs.float(),0)
#last and repeat is to check whether it shoud stop
last=0
repeat=0
for i in range(50):
outputs=NN(X,Y,random_effect)
epochs=5
for e in range(epochs):
Random_effect_true=Y-outputs.data.numpy()
mu,Sigma1=E_step(sigma_u,sigma_e,Random_effect_true,Z)
sigma_u,sigma_e=M_step(mu,Sigma1,Random_effect_true,Z)
random_effect=np.zeros((len(X),1))
for j in range(len(Z)):
z=Z[j]
p=np.dot(z,mu)
random_effect[j]=p
random_effect_test=np.zeros((len(z_test),1))
for j in range(len(z_test)):
z=z_test[j]
p=np.dot(z,mu)
random_effect_test[j]=p
inputs_test=pt.from_numpy(x_test)
labels_test=pt.from_numpy(y_test)
inputs_test = pt.autograd.Variable(inputs_test)
labels_test = pt.autograd.Variable(labels_test)
random_e_test=pt.from_numpy(random_effect_test).float()
outputs_test = model(inputs_test.float(),random_e_test)
error_percentage=AccuarcyCompute(outputs_test,labels_test.float())
if error_percentage[2]==last:
repeat+=1
else:
last=error_percentage[2]
if repeat>5:
break
print("Testing",error_percentage)
if __name__ == '__main__':
Xdata,y=gen_data(5,size=1000)
Xdata=np.concatenate((np.ones((Xdata.shape[0],1)),Xdata),axis=1).astype(float)
for i in range(len(y)):
EMNN(y[i])
'''
Testing (0.0015800932, 0.029433494, 4.1)
Testing (0.0029981958, 0.044504516, 4.91)
Testing (0.002603048, 0.0364661, 4.62)
Testing (0.002724695, 0.042232957, 6.26)
Testing (0.002901133, 0.038665723, 4.27)
Testing (0.0019755429, 0.036214557, 3.75)
Testing (0.002415244, 0.040518273, 6.68)
Testing (0.0023834745, 0.035962757, 3.54)
Testing (0.0021507577, 0.035938308, 3.04)
Testing (0.0031414144, 0.044444475, 3.79)
N=10000
Testing (0.0024926742, 0.040062033, 5.71)
Testing (0.0029507317, 0.043557882, 6.11)
Testing (0.003065372, 0.04415746, 5.99)
Testing (0.0030026636, 0.04395273, 5.34)
Testing (0.0030103307, 0.043855403, 4.83)
N=1000
Testing (0.002531604, 0.038599867, 5.47)
Testing (0.0030081624, 0.043093387, 6.01)
Testing (0.0031148489, 0.044860844, 6.06)
Testing (0.0031499858, 0.044847127, 5.64)
Testing (0.002765486, 0.042059153, 4.28)
N=3000
Testing (0.0025788886, 0.040610652, 5.65)
Testing (0.0029482488, 0.04301046, 6.08)
Testing (0.0030424101, 0.043871272, 6.07)
Testing (0.0027394677, 0.04155968, 5.01)
Testing (0.0029647427, 0.043070924, 5.02)
N=5000
Testing (0.0025914374, 0.040200982, 5.57)
Testing (0.002770738, 0.04197966, 5.88)
Testing (0.0029406706, 0.04297577, 5.7)
Testing (0.0030163485, 0.043473613, 5.17)
Testing (0.0031185846, 0.04446688, 4.84)
N=8000
Testing (0.0024359503, 0.03923105, 5.52)
Testing (0.0028627987, 0.042770088, 5.95)
Testing (0.0029198667, 0.04259068, 5.69)
Testing (0.0030189105, 0.043778256, 5.22)
Testing (0.0030523143, 0.043963313, 5.03)
'''