-
Notifications
You must be signed in to change notification settings - Fork 7
/
vae.py
241 lines (207 loc) · 8.39 KB
/
vae.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
import tensorflow as tf
import numpy as np
import pickle as pkl
import os
from scipy.stats import gaussian_kde
from tf_utils import logger, gpu_session, clear_dir
tfd = tf.contrib.distributions
class VariationalAutoEncoder(object):
"""
Variational Auto Encoder
"""
def __init__(self, encoder, decoder, datasets, optimizer, logdir):
self.encoder = encoder
self.decoder = decoder
self.datasets = datasets
self.optimizer = optimizer
self.logdir = logdir
self._create_datasets()
self._create_loss()
self._create_optimizer(encoder, decoder, optimizer)
self._create_summary()
self._create_evaluation(encoder, decoder)
self._create_session(logdir)
logger.configure(logdir, format_strs=['stdout', 'log'])
def _create_datasets(self):
datasets = self.datasets
self.iterator = iterator = tf.data.Iterator.from_structure(
output_types=datasets.train.output_types, output_shapes=datasets.train.output_shapes
)
self.train_init = iterator.make_initializer(datasets.train)
self.test_init = iterator.make_initializer(datasets.test)
def _create_loss(self):
self.x = self.iterator.get_next()[0]
z, logqzx = self.encoder.sample_and_log_prob(self.x)
x_, logpxz, logpz = self.decoder.sample_and_log_prob(z, self.x)
self.z = z
self.encoder_loss = logqzx - logpz
self.decoder_loss = -logpxz
self.nll = tf.reduce_mean(self.decoder_loss)
self.elbo = tf.reduce_mean(self.encoder_loss)
self.loss = self.nll + self.elbo
def _create_optimizer(self, encoder, decoder, optimizer):
encoder_grads_and_vars = optimizer.compute_gradients(self.loss, encoder.vars)
decoder_grads_and_vars = optimizer.compute_gradients(self.loss, decoder.vars)
self.trainer = tf.group(optimizer.apply_gradients(encoder_grads_and_vars),
optimizer.apply_gradients(decoder_grads_and_vars))
def _create_summary(self):
with tf.name_scope('train'):
self.train_summary = tf.summary.merge([
tf.summary.scalar('elbo', self.elbo),
tf.summary.scalar('nll', self.nll),
tf.summary.scalar('loss', self.loss)
])
def _create_evaluation(self, encoder, decoder):
x = self.iterator.get_next()[0]
self.z_mi = encoder.sample_and_log_prob(x)[0]
self.log_q_z_x = encoder.sample_and_log_prob(x)[1]
def _create_session(self, logdir):
self.summary_writer = tf.summary.FileWriter(logdir=logdir)
self.sess = gpu_session()
self.saver = tf.train.Saver()
self.logdir = logdir
def _update_optimizer(self):
pass
def _debug(self):
pass
def _train(self):
self._debug()
self.sess.run([self.trainer])
def _log(self, it):
if it % 10 == 0:
loss, nll, elbo = self.sess.run([self.loss, self.nll, self.elbo])
logger.log("Iteration %d: loss %.4f nll %.4f elbo %.4f" % (it, loss, nll, elbo))
self.summary_writer.add_summary(self.sess.run(self.train_summary), it)
def train(self, num_epochs, num_iters=None):
self.sess.run(tf.global_variables_initializer())
it = 0
for epoch in range(num_epochs):
self.sess.run(self.train_init)
self._update_optimizer()
while True:
try:
self._train()
it += 1
self._log(it)
except tf.errors.OutOfRangeError:
break
if num_iters and it > num_iters:
break
if epoch % 100 == 1:
print('Saving to: ', os.path.join(self.logdir, 'model/model.ckpt'))
self.saver.save(sess=self.sess, save_path=os.path.join(self.logdir, 'model/model.ckpt'))
self.sess.run(self.train_init)
print('Saving to: ', os.path.join(self.logdir, 'model/model.ckpt'))
self.saver.save(sess=self.sess, save_path=os.path.join(self.logdir, 'model/model.ckpt'))
def test(self):
print('Loading from ', os.path.join(self.logdir, 'model/model.ckpt'))
self.saver.restore(sess=self.sess, save_path=os.path.join(self.logdir, 'model/model.ckpt'))
self.sess.run(self.train_init)
z_mis = []
while True:
try:
z_mi = self.sess.run(self.z_mi)
z_mis.append(z_mi)
except tf.errors.OutOfRangeError:
break
z_mis = np.concatenate(z_mis, axis=0)
kde = gaussian_kde(z_mis.transpose())
self._evaluate_over_test_set(
[self.elbo, self.nll, self.log_q_z_x],
['elbo', 'nll', 'logqzx']
)
self._estimate_mutual_information_continuous(self.z_mi, self.log_q_z_x)
def _evaluate_over_test_set(self, keys, strs):
self.sess.run(self.test_init)
d = {'test_' + s: [] for s in strs}
while True:
try:
ks = self.sess.run(keys)
for i in range(len(keys)):
d['test_' + strs[i]].append(ks[i])
except tf.errors.OutOfRangeError:
break
for k in d.keys():
d[k] = np.mean(d[k])
self._write_evaluation(d)
def _evaluate_over_train_set(self, keys, strs):
self.sess.run(self.train_init)
d = {'train_' + s: [] for s in strs}
while True:
try:
ks = self.sess.run(keys)
# import ipdb; ipdb.set_trace()
for i in range(len(keys)):
d['train_' + strs[i]].append(ks[i])
except tf.errors.OutOfRangeError:
break
for k in d.keys():
d[k] = np.mean(d[k])
self._write_evaluation(d)
def _estimate_mutual_information_continuous(self, z_var, qzx_var, label='qzx'):
self.sess.run(self.train_init)
zs, mis = [], []
while True:
try:
z_mi = self.sess.run(z_var)
zs.append(z_mi)
except tf.errors.OutOfRangeError:
break
zs = np.concatenate(zs, axis=0)
kde = gaussian_kde(zs.transpose())
self.sess.run(self.test_init)
while True:
try:
z, lqzx = self.sess.run([z_var, qzx_var])
mi = lqzx - kde.logpdf(z.transpose())
mis.append(np.mean(mi))
except tf.errors.OutOfRangeError:
break
d = {'mi_' + label: np.mean(mis)}
self._write_evaluation(d)
def _estimate_mutual_information_discrete(self, z_var, qzx_var, qu0_var, qu1_var, y_var, label='quz'):
self.sess.run(self.train_init)
zs, mis = [], []
mis0, mis1 = [], []
zs0, zs1 = [], []
while True:
try:
z_mi, y_mi = self.sess.run([z_var, y_var])
zs.append(z_mi)
zs0.append(z_mi[np.where(y_mi == 0)])
zs1.append(z_mi[np.where(y_mi == 1)])
except tf.errors.OutOfRangeError:
break
zs = np.mean(np.concatenate(zs, axis=0), axis=0)
self.sess.run(self.test_init)
while True:
try:
z, lqzx, qu0, qu1, y = self.sess.run([z_var, qzx_var, qu0_var, qu1_var, y_var])
mi = lqzx - (z * np.log(zs) + (1.0 - z) * np.log(1.0 - zs))
mis.extend(mi)
mi = qu0 - (z * np.log(zs) + (1.0 - z) * np.log(1.0 - zs))
mis0.extend(mi[np.where(y == 0)])
mi = qu1 - (z * np.log(zs) + (1.0 - z) * np.log(1.0 - zs))
mis1.extend(mi[np.where(y == 1)])
except tf.errors.OutOfRangeError:
break
d = {
'mi_' + label: np.mean(mis),
'mi0_' + label: np.mean(mis0),
'mi1_' + label: np.mean(mis1),
'mi01_' + label: np.mean(mis0 + mis1)
}
self._write_evaluation(d)
def _write_evaluation(self, d):
logger.logkvs(d)
logger.dumpkvs()
try:
with open(os.path.join(self.logdir, 'eval.pkl'), 'rb') as f:
d_ = pkl.load(f)
except FileNotFoundError:
d_ = {}
for k in d_.keys():
if k not in d:
d[k] = d_[k]
with open(os.path.join(self.logdir, 'eval.pkl'), 'wb') as f:
pkl.dump(d, f)