-
Notifications
You must be signed in to change notification settings - Fork 14
/
SIVAE.py
246 lines (180 loc) · 7.65 KB
/
SIVAE.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
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import matplotlib
matplotlib.use('Agg')
import numpy as np
import os
import sys
import seaborn as sns
import scipy.spatial.distance
from matplotlib import pyplot as plt
import pandas as pd
import scipy.stats as stats
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
slim=tf.contrib.slim
Exponential=tf.contrib.distributions.Exponential(rate=1.0)
Normal=tf.contrib.distributions.Normal(loc=0., scale=1.)
Mvn=tf.contrib.distributions.MultivariateNormalDiag
Bernoulli = tf.contrib.distributions.Bernoulli
plt.ioff()
sys.path.append(os.getcwd())
#%%
def sample_psi(x,noise_dim,K,z_dim,reuse=False):
with tf.variable_scope("hyper_psi") as scope:
if reuse:
scope.reuse_variables()
x_0 = tf.expand_dims(x,axis=1)
x_1 = tf.tile(x_0,[1,K,1]) #N*K*784
B3 = Bernoulli(0.5)
e3 = tf.cast(B3.sample([tf.shape(x)[0],K,noise_dim[0]]),tf.float32)
input_ = tf.concat([e3,x_1],axis=2)
h3 = slim.stack(input_,slim.fully_connected,[500,500,noise_dim[0]])
B2 = Bernoulli(0.5)
e2 = tf.cast(B2.sample([tf.shape(x)[0],K,noise_dim[1]]),tf.float32)
input_1 = tf.concat([h3,e2,x_1],axis=2)
h2 = slim.stack(input_1,slim.fully_connected,[500,500,noise_dim[1]])
B1 = Bernoulli(0.5)
e1 = tf.cast(B1.sample([tf.shape(x)[0],K,noise_dim[2]]),tf.float32)
h1 = slim.stack(tf.concat([h2,e1,x_1],axis=2),slim.fully_connected,[500,500,500])
mu = tf.reshape(slim.fully_connected(h1,z_dim,activation_fn=None,scope='implicit_hyper_mu'),[-1,K,z_dim])
return mu
def sample_logv(x,noise_dim,z_dim,reuse=False):
with tf.variable_scope("hyper_sigma") as scope:
if reuse:
scope.reuse_variables()
net1 = slim.stack(x,slim.fully_connected,[500,500],scope='sigma')
z_logv = tf.reshape(slim.fully_connected(net1,z_dim,activation_fn=None,scope='sigma2'),[-1,z_dim])
return z_logv
def sample_n(psi,sigma):
eps = tf.random_normal(shape=tf.shape(psi))
z=psi+eps*sigma
return z
def decoder(z,x_dim,reuse=False):
with tf.variable_scope("decoder") as scope:
if reuse:
scope.reuse_variables()
net3 = slim.stack(z,slim.fully_connected,[500,500,500],scope='decoder_1')
logits_x = slim.fully_connected(net3,x_dim,activation_fn=None,scope='decoder_2')
return logits_x
#%%
tf.reset_default_graph()
z_dim = 64
noise_dim = [150,100,50]
x_dim = 784
eps = 1e-10
WU = tf.placeholder(tf.float32, shape=()) #warm-up
x = tf.placeholder(tf.float32,[None,x_dim])
J = tf.placeholder(tf.int32, shape=()) #estimate h
merge = tf.placeholder(tf.int32, shape=[])
K = tf.placeholder(tf.int32, shape=()) #iwae
z_logv = sample_logv(x,noise_dim,z_dim)
z_logv_iw = tf.tile(tf.expand_dims(z_logv,axis=1),[1,K,1])
sigma_iw1 = tf.exp(z_logv_iw/2)
sigma_iw2 = tf.cond(merge>0,lambda:tf.tile(tf.expand_dims(sigma_iw1,axis=2),[1,1,J+1,1]),
lambda:tf.tile(tf.expand_dims(sigma_iw1,axis=2),[1,1,J,1]))
psi_iw = sample_psi(x,noise_dim,K,z_dim)
z_sample_iw = sample_n(psi_iw,sigma_iw1)
z_sample_iw1 = tf.expand_dims(z_sample_iw,axis=2)
z_sample_iw2 = tf.cond(merge>0,lambda:tf.tile(z_sample_iw1,[1,1,J+1,1]),
lambda:tf.tile(z_sample_iw1,[1,1,J,1]))
psi_iw_star = sample_psi(x,noise_dim,J,z_dim,reuse=True)
psi_iw_star0 = tf.expand_dims(psi_iw_star,axis=1)
psi_iw_star1 = tf.tile(psi_iw_star0,[1,K,1,1])
psi_iw_star2 = tf.cond(merge>0,lambda:tf.concat([psi_iw_star1, tf.expand_dims(psi_iw,axis=2)],2),
lambda:psi_iw_star1)
ker = tf.exp(-0.5*tf.reduce_sum(tf.square(z_sample_iw2-psi_iw_star2)/tf.square(sigma_iw2+eps),3))
log_H_iw = tf.log(tf.reduce_mean(ker,axis=2))-0.5*tf.reduce_sum(z_logv_iw,2) #change to tf.reduce_logsumexp if there is NA
log_prior_iw = -0.5*tf.reduce_sum(tf.square(z_sample_iw),2)
x_iw = tf.tile(tf.expand_dims(x,axis=1),[1,K,1])
logits_x_iw = decoder(z_sample_iw,x_dim)
p_x_iw = Bernoulli(logits=logits_x_iw)
reconstruct_iw = p_x_iw.mean()
log_lik_iw = tf.reduce_sum( x_iw * tf.log(reconstruct_iw + eps)
+ (1-x_iw) * tf.log(1 - reconstruct_iw + eps),2)
loss_iw0 = -tf.reduce_logsumexp(log_lik_iw+(log_prior_iw-log_H_iw)*WU,1)+tf.log(tf.cast(K,tf.float32))
loss_iw = tf.reduce_mean(loss_iw0)
var_all = slim.get_model_variables()
lr=tf.constant(0.001)
g_step = tf.Variable(0, trainable=False)
train_op = tf.train.AdamOptimizer(learning_rate=lr).minimize(loss_iw,var_list=var_all,global_step=g_step)
init_op=tf.global_variables_initializer()
#%%
mnist = input_data.read_data_sets(os.getcwd()+'/MNIST', one_hot=True)
train_data=mnist.train
test_data=mnist.test
dat_train=[]
dat_test=[]
sess=tf.InteractiveSession()
sess.run(init_op)
print("This is SIVAE_MNIST test")
training_epochs=2000
batch_size = 200
total_points = mnist.train.num_examples
total_batch = int(total_points / batch_size)
display_step=1
total_test_batch = int(mnist.test.num_examples / batch_size)
J_value = 1
warm_up = 0
from time import sleep
for epoch in range(training_epochs):
avg_cost = 0.
avg_cost_test = 0.
np_lr = 0.001 * 0.75**(epoch/100)
warm_up = np.min([epoch/300,1])
if epoch<1900:
J_value = 1
else:
J_value = 50
for i in range(total_batch):
train_xs_0,_ = train_data.next_batch(batch_size)
train_xs = np.random.binomial(1,train_xs_0)
_ = sess.run([train_op],{x:train_xs,lr:np_lr,merge:1,J:J_value,K:1,WU:warm_up})
if epoch>1900:
for k in range(total_batch):
train_xs_0,_ = train_data.next_batch(batch_size)
train_xs = np.random.binomial(1,train_xs_0)
cost=sess.run(loss_iw,{x:train_xs,J:J_value,merge:1,K:1,WU:1.0})
avg_cost += cost / total_batch
for j in range(total_test_batch):
test_xs_0,_ = test_data.next_batch(batch_size)
test_xs = np.random.binomial(1,test_xs_0)
cost_test=sess.run(loss_iw,{x:test_xs,J:J_value,merge:1,K:1,WU:1.0})
avg_cost_test += cost_test / total_test_batch
dat_train.append([epoch,avg_cost])
dat_test.append([epoch,avg_cost_test])
if epoch % display_step == 0:
print("Epoch:", '%04d' % epoch,
"cost_train=", "{:.9f}".format(avg_cost),
"cost_test=", "{:.9f}".format(avg_cost_test))
avg_evi_test = 0
for j in range(total_test_batch):
test_xs_0,_ = test_data.next_batch(batch_size)
test_xs = np.random.binomial(1,test_xs_0)
evi_test=sess.run(loss_iw,{x:test_xs,J:J_value,merge:1,K:1000,WU:1.0})
avg_evi_test += evi_test / total_test_batch
L_1000 = avg_evi_test
print("&&&&&&&& The final test evidence is", L_1000)
if not os.path.exists('out/'):
os.makedirs('out/')
dat0 = np.array(dat_train)
dat1 = np.array(dat_test)
df0 = pd.DataFrame({'epoch':dat0[:,0],'train':dat0[:,1]})
df1 = pd.DataFrame({'epoch':dat1[:,0],'test':dat1[:,1]})
df = pd.concat([df0,df1], ignore_index=True, axis=1)
name_data1 = 'out/data_dim4_'+str(noise_dim)+'.csv'
df.to_csv(name_data1,index=False)
name_fig1 = 'out/slim_ELBO_dim4_'+str(noise_dim)+'.png'
if 1:
plt.figure()
dat0 = np.array(dat_train)
dat1 = np.array(dat_test)
plt.plot(dat0[:,0],dat0[:,1],'o-',label='train')
plt.plot(dat1[:,0],dat1[:,1],'o-',label='test')
plt.legend(fontsize = 'x-large')
plt.title("Training performance",fontsize = 'x-large')
plt.ylabel('nats',fontsize = 'x-large')
plt.xlabel('epoch',fontsize = 'x-large')
plt.savefig(name_fig1, bbox_inches='tight')
plt.close('all')