MMD/Example_VAE.py at master · eugenium/MMD · GitHub

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
"""
Authors:
Eugene Belilovsky
"""

"""This script trains an auto-encoder on the MNIST dataset and keeps track of the lowerbound"""

#python trainmnist.py -s mnist.npy

import VariationalAutoencoder
import numpy as np
import scipy as sp
import time,os
import gzip, cPickle,copy,pickle
from sklearn import linear_model
from sklearn.cross_validation import train_test_split
from mmd import MMD_3_Sample_Test

print "Loading MNIST data"
#Retrieved from: http://deeplearning.net/data/mnist/mnist.pkl.gz

f = gzip.open('mnist.pkl.gz', 'rb')
(x_train, t_train), (x_valid, t_valid), (x_test, t_test)  = cPickle.load(f)
f.close()
x_train=(x_train>0).astype('float')
x_valid=(x_valid>0).astype('float')
x_test=(x_test>0).astype('float')

data=x_train
samp_size=1000


verbose=True

runs=25
dimZ = 20
HU_decoder = 400
HU_encoder = HU_decoder

dimZ2 = dimZ
HU_decoder2 = HU_decoder
HU_encoder2 = HU_decoder2

batch_size = 100
L = 1
learning_rate = 0.01

sig=0.05

ratio=0.3
#ratios=np.array([0.5,1,2])
t_size2=2000
t_size1=int(t_size2/ratio)

set1,set2=train_test_split(range(data.shape[0]),train_size=t_size1+t_size2)
data1 = data[set1[0:t_size1],:]
data2 = data[set1[t_size1:t_size2+t_size1],:]
set2=set2[0:samp_size]
data_holdout=data[set2,:]


[N1,dimX] = data1.shape
[N2,dimX] = data2.shape
encoder1 = VariationalAutoencoder.VA(HU_decoder,HU_encoder,dimX,dimZ,batch_size,L,learning_rate,continous=False)
encoder2 = VariationalAutoencoder.VA(HU_decoder2,HU_encoder2,dimX,dimZ2,batch_size,L,learning_rate,continous=False)


print "Creating Theano functions"
encoder1.createGradientFunctions()
encoder2.createGradientFunctions()
print "Initializing weights and biases"
encoder1.initParams()
encoder2.initParams()

begin = time.time()
maxiter=2000
testlowerbound1=testlowerbound2=-np.Inf
for j in xrange(maxiter):
    encoder1.iterate(data1)
    if j%1 == 0:
        oldlower=testlowerbound1
        train_lower1=encoder1.getLowerBound(data1)
        testlowerbound1 = encoder1.getLowerBound(data_holdout)
        if(verbose):
            print("Encoder 1 Iteration %d| lower bound train = %.2f |lower bound test 1= %.2f"
                  % (j, train_lower1/float(N1),testlowerbound1/samp_size))
        if(oldlower>=testlowerbound1):
            break
        best_encoder1=copy.deepcopy(encoder1)


encoder1=best_encoder1

for j in xrange(maxiter):
    encoder2.iterate(data2)

    if j%1 == 0:
        oldlower=testlowerbound2
        train_lower2=encoder2.getLowerBound(data2)
        testlowerbound2 = encoder2.getLowerBound(data_holdout)
        if(verbose):
            print("Encoder 2 Iteration %d| lower bound train = %.2f |lower bound test 1= %.2f"
                  % (j, train_lower2/float(N2),testlowerbound2/samp_size))
        if(oldlower>testlowerbound2):
            break
        best_encoder2=copy.deepcopy(encoder2)


encoder2=best_encoder2
end=time.time()


samples1=encoder1.sample(N=samp_size)
samples2=encoder2.sample(N=samp_size)

pvalue,tstat,sigma,MMDXY,MMDXZ=MMD_3_Sample_Test(data_holdout,samples1,samples2,computeMMDs=True)
print("MMD(enc1 samples,real): %.4f MMD(enc2 samples,real): %.4f , pvalue: %.2f"%(MMDXY,MMDXZ,pvalue))
#Regressions
##Train regression

data1_enc=encoder1.encode(x_valid)
data2_enc=encoder2.encode(x_valid)
test1_enc=encoder1.encode(x_test)
test2_enc=encoder2.encode(x_test)

LogReg_VAE = linear_model.LogisticRegression()
LogReg_VAE.fit(data1_enc,t_valid)
vae1_score = LogReg_VAE.score(test1_enc, t_test)
LogReg_VAE = linear_model.LogisticRegression()
LogReg_VAE.fit(data2_enc,t_valid)
vae2_score = LogReg_VAE.score(test2_enc, t_test)
print("Accuracy 1:%.2f Accuracy2:%.2f"%(vae1_score,vae2_score))