-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathvae.py
183 lines (153 loc) · 8.51 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
import tensorflow as tf
import numpy as np
from PIL import Image
import random
import glob
import os.path
import sys
# Model
class VAEGANModel:
def __init__(self, latent_size=1024, mb_size=24, kappa=1.0, gamma=20.0):
def vaegan_conv(inp, filters, filter_size, name, activation=tf.nn.relu):
return activation(tf.layers.batch_normalization(
tf.layers.conv2d(inp,
filters,
filter_size,
strides=1,
padding='same',
activation=None,
name=name), training=True))
def vaegan_deconv(inp, filters, filter_size, stride, name, activation=tf.nn.relu):
return activation(tf.layers.batch_normalization(
tf.contrib.layers.convolution2d_transpose(inp,
num_outputs=filters,
kernel_size=filter_size,
stride=stride,
padding='same',
activation_fn=None,
scope=name), training=True))
def vaegan_dense(inp, hus, name, activation=tf.nn.relu):
return activation(tf.layers.batch_normalization(
tf.layers.dense(inp, hus, activation=None, name=name), training=True))
X_raw = tf.placeholder(tf.float32, shape=(mb_size, 64, 64, 3))
X = X_raw / 225.0
# Encoder
conv1 = vaegan_conv(X, 64, 5, "enc1")
conv2 = vaegan_conv(conv1, 128, 3, "enc2")
conv3 = vaegan_conv(conv2, 256, 3, "enc3")
flt = tf.contrib.layers.flatten(conv3)
enc_d1 = vaegan_dense(flt, 128, "enc4")
mu_sig = vaegan_dense(enc_d1, latent_size * 2, "enc5")
mu = mu_sig[:, :latent_size]
sigma = mu_sig[:, latent_size:]
epsilon = tf.random_normal([mb_size, latent_size], 0.0, 1.0, dtype=tf.float32)
z = mu + (tf.sqrt(tf.exp(sigma)) * epsilon)
p_z = tf.random_normal([mb_size, latent_size], 0.0, 1.0, dtype=tf.float32)
z_comb = tf.concat([z, p_z], 0)
# Decoder / Reconstructor
dec_d1 = vaegan_dense(z_comb, 256, "dec1")
dec_d1_rshp = tf.reshape(dec_d1, [-1, 16, 16, 1])
deconv_1 = vaegan_deconv(dec_d1_rshp, 256, 3, 1, "dec2")
deconv_2 = vaegan_deconv(deconv_1, 128, 5, 1, "dec3")
deconv_3 = vaegan_deconv(deconv_2, 32, 5, 2, "dec4")
deconv_4 = vaegan_deconv(deconv_3, 3, 5, 2, "dec5", activation=tf.nn.sigmoid)
X_hat = deconv_4[:mb_size]
# GAN Discriminator
X_gan = tf.concat([X, deconv_4], 0)
gan_conv1 = tf.layers.conv2d(X_gan, 32, 5, 1, name="gan1", activation=tf.nn.relu) # No BN
gan_conv2 = vaegan_conv(gan_conv1, 128, 5, "gan2")
gan_conv3 = vaegan_conv(gan_conv2, 256, 5, "gan3")
gan_conv4 = vaegan_conv(gan_conv3, 256, 5, "gan4")
gan_dense1 = vaegan_dense(gan_conv4, 512, "gan5")
gan_out = vaegan_dense(gan_dense1, 1, "gan6", activation=tf.nn.sigmoid)
# Loss
reconstruction_loss = -1.0 * tf.reduce_sum(X * tf.log(tf.clip_by_value(X_hat, 1e-10, 1.0))
+ (1.0 - X) * tf.log(tf.clip_by_value(1.0 - X_hat, 1e-10, 1.0)))
KL_divergence = -1.0 * tf.reduce_sum(1.0 + sigma - mu**2 - tf.exp(sigma))
GAN_X = gan_out[:mb_size]
GAN_X_hat = gan_out[mb_size:mb_size*2]
GAN_pz = gan_out[mb_size*2:]
GAN_loss = -1.0 * tf.reduce_sum(tf.log(tf.clip_by_value(GAN_X, 1e-10, 1.0))
+ tf.log(tf.clip_by_value(1.0 - GAN_X_hat, 1e-10, 1.0))
+ tf.log(tf.clip_by_value(1.0 - GAN_pz, 1e-10, 1.0)))
enc_loss = tf.reduce_mean(reconstruction_loss + kappa * KL_divergence)
dec_loss = tf.reduce_mean(gamma * reconstruction_loss - GAN_loss)
gan_loss = tf.reduce_mean(GAN_loss)
# Define subnets
enc = filter(lambda x: x.name.startswith("enc"), tf.trainable_variables())
dec = filter(lambda x: x.name.startswith("dec"), tf.trainable_variables())
gan = filter(lambda x: x.name.startswith("gan"), tf.trainable_variables())
# Opt
train_step_enc = tf.train.AdamOptimizer(0.001).minimize(enc_loss, var_list=enc)
train_step_dec = tf.train.AdamOptimizer(0.001).minimize(dec_loss, var_list=dec)
train_step_gan = tf.train.AdamOptimizer(0.001).minimize(gan_loss, var_list=gan)
# Summary (Tensorboard)
tf.summary.scalar("enc_loss", enc_loss)
tf.summary.scalar("dec_loss", dec_loss)
tf.summary.scalar("gan_loss", gan_loss)
tf.summary.scalar("reconstruction_loss", reconstruction_loss)
tf.summary.scalar("kl_divergence", KL_divergence)
tf.summary.scalar("GAN_loss", GAN_loss)
tf.summary.histogram("z", z)
tf.summary.histogram("gan_out", gan_out)
tf.summary.histogram("gan_X", GAN_X)
tf.summary.histogram("gan_X_hat", GAN_X_hat)
tf.summary.image("in", X, max_outputs=1)
tf.summary.image("out", X_hat, max_outputs=1)
summary_op = tf.summary.merge_all()
# Model I/O
saver = tf.train.Saver()
# Outbound
self.saver = saver
self.summary_op = summary_op
self.train_step_enc = train_step_enc
self.train_step_dec = train_step_dec
self.train_step_gan = train_step_gan
self.reconstruction_loss = reconstruction_loss
self.KL_divergence = KL_divergence
self.X_hat = X_hat
self.X_raw = X_raw
self.z = z
self.mu = mu
class Batcher():
def __init__(self, root_path):
self.paths = glob.glob("{}/*.jpg".format(root_path))
self.i = 0
def batch(self, mb_size):
b = np.zeros((mb_size, 64, 64, 3))
for i in xrange(mb_size):
img = Image.open(self.paths[self.i])
img.load()
b[i] = np.asarray(img)
self.i += 1
if self.i >= len(self.paths):
self.i = 0
random.shuffle(self.paths)
return b
if __name__ == "__main__":
with tf.Session() as sess:
model = VAEGANModel(mb_size=32)
batcher = Batcher("data")
# Summary
summary_writer = tf.summary.FileWriter('logs/vaegan', graph=sess.graph)
# Load / Init model weights
if os.path.isfile("models/vaegan.ckpt.meta"):
print "* Restoring saved parameters"
model.saver.restore(sess, "models/vaegan.ckpt")
else:
print "* Initializing parameters"
sess.run(tf.global_variables_initializer())
print "# Logging step and loss, see Tensorboard for more information..."
for epoch in xrange(100):
for mb_n in xrange(int(202597/32)):
mb = batcher.batch(32)
_, _, _, summary = sess.run([ model.train_step_enc,
model.train_step_dec,
model.train_step_gan,
model.summary_op], feed_dict={model.X_raw: mb})
if mb_n % 3 == 0:
summary_writer.add_summary(summary, mb_n)
print "MB {} of {} of epoch {}".format(mb_n, int(202597/8), epoch)
if mb_n % 250 == 0:
print "* Saving model"
model.saver.save(sess, "models/vaegan.ckpt")