forked from cryptexis/sgan-tensorflow
-
Notifications
You must be signed in to change notification settings - Fork 0
/
sgan.py
286 lines (229 loc) · 10.9 KB
/
sgan.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
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
import os
import time
import numpy as np
import re
from utils.data_util import get_data, save_tensor
from model.ops import *
class SpatialGan(object):
real_images = None
random_noise = None
G_net = None
G_loss = None
G_vars = None
D_net_real = None
D_net_fake = None
D_logits_real = None
D_logits_fake = None
D_loss_real = None
D_loss_fake = None
D_loss = None
D_vars = None
lr_sum = None
lf_sum = None
lg_sum = None
lt_sum = None
img_sum = None
hyper_params = None
saver = None
def __init__(self, sess):
self.opt = tf.app.flags.FLAGS
self.sess = sess
self.batch_size = self.opt.batch_size
self.image_size = (2**self.opt.num_layers)*(self.opt.z_size - 1) + 1
# define the dimensionality of the input tensor for discriminator network
self.image_shape = [self.image_size, self.image_size, 3]
# define the dimensionality of the input tensor for generator network
self.z_dim = [self.opt.z_size, self.opt.z_size, self.opt.z_dim]
# depending on the number of layers define the depth of each layer D_network
self.d_layers = []
self.d_filters = [2**(n+6) for n in range(self.opt.num_layers-1)]+[1]
self.d_weights = []
# same for the G_network
self.g_layers = []
self.g_weights = []
self.g_filters = [3] + [2 ** (n + 6) for n in range(self.opt.num_layers - 1)]
self.g_filters = self.g_filters[::-1]
self.d_bn_layers = []
self.g_bn_layers = []
# batch normalization layers
for i in range(0, self.opt.num_layers - 2):
self.d_bn_layers.append(BatchNorm(name='d_bn'+str(i+1)))
for i in range(0, self.opt.num_layers - 1):
self.g_bn_layers.append(BatchNorm(name='g_bn'+str(i+1)))
def build_model(self):
"""
Model and Loss definitions
:return:
"""
self.real_images = tf.placeholder(tf.float32,
[self.batch_size] + self.image_shape,
name='real_images')
self.random_noise = tf.placeholder(tf.float32,
[self.batch_size] + self.z_dim,
name='z')
# forward pass through generator
self.G_net = self.generator(self.random_noise)
# forward pass through discriminator / real images
self.D_net_real, self.D_logits_real = self.discriminator(self.real_images)
# forward pass through discriminator / generated images + reuse variables from the first network
self.D_net_fake, self.D_logits_fake = self.discriminator(self.G_net, reuse=True)
if self.opt.is_train:
# discriminative loss on data points sampled from data distribution
self.D_loss_real = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=self.D_logits_real,
labels=tf.ones_like(self.D_logits_real)))
self.lr_sum = tf.summary.scalar("D_loss_real", self.D_loss_real)
# discriminative loss on data points sampled from generated distribution
self.D_loss_fake = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=self.D_logits_fake,
labels=tf.zeros_like(self.D_logits_fake)))
self.lf_sum = tf.summary.scalar("D_loss_fake", self.D_loss_fake)
# generative loss
self.G_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=self.D_logits_fake,
labels=tf.ones_like(self.D_logits_fake)))
self.lg_sum = tf.summary.scalar("G_loss", self.G_loss)
self.img_sum = tf.summary.image("output", self.G_net, 3)
# Discriminative loss + L2 decay on weights
self.D_loss = self.D_loss_real + self.D_loss_fake
self.lt_sum = tf.summary.scalar("D_loss", self.D_loss)
t_vars = tf.trainable_variables()
self.D_vars = [var for var in t_vars if 'd_' in var.name]
self.G_vars = [var for var in t_vars if 'g_' in var.name]
self.saver = tf.train.Saver(reshape=True, max_to_keep=self.opt.max_iterations)
def train(self):
"""
Training steps for Spatial GAN
:return:
"""
# define optimizer for both networks
d_optim = tf.train.AdamOptimizer(self.opt.learning_rate, beta1=self.opt.beta1) \
.minimize(self.D_loss, var_list=self.D_vars)
g_optim = tf.train.AdamOptimizer(self.opt.learning_rate, beta1=self.opt.beta1) \
.minimize(self.G_loss, var_list=self.G_vars)
tf.global_variables_initializer().run()
g_sum = tf.summary.merge([self.lg_sum, self.img_sum])
d_sum = tf.summary.merge([self.lr_sum, self.lf_sum, self.lt_sum])
sample_image = np.random.uniform(-1, 1, [self.opt.batch_size]+self.z_dim).astype(np.float32)
writer = tf.summary.FileWriter(self.opt.checkpoint_dir, self.sess.graph)
start_time = time.time()
epoch = 1
while epoch < self.opt.max_iterations:
batch_images = get_data(self.opt.dataset, self.image_size, self.batch_size)
batch_z = np.random.uniform(-1, 1, [self.opt.batch_size]+self.z_dim).astype(np.float32)
if epoch % 2 != 0:
print "Update D:"
# Update D network
_, summary_str = self.sess.run([d_optim, d_sum],
feed_dict={
self.real_images: batch_images,
self.random_noise: batch_z
})
else:
print "Update G:"
# Update G network
_, summary_str = self.sess.run([g_optim, g_sum],
feed_dict={
self.random_noise: batch_z
})
if epoch % self.opt.checkpoint_interval == 0:
writer.add_summary(summary_str, epoch)
img_out = self.G_net.eval({self.random_noise: sample_image})
save_tensor(img_out[0], os.path.join(self.opt.sample_dir, str(epoch)+'.jpg'))
if epoch % 100 == 0:
self.saver.save(self.sess, os.path.join(self.opt.checkpoint_dir, 'fck.ckpt'), global_step=epoch)
time_delta = time.time() - start_time
print("Epoch: [%2d] time: %4.4f" % (epoch, time_delta))
epoch += 1
def discriminator(self, image, reuse=False):
"""
In discriminator network batch normalization
used on all layers except input and output
:param image:
:param reuse:
:return:
"""
with tf.variable_scope("discriminator") as scope:
if reuse:
scope.reuse_variables()
# down sample the first layer
conv, weight = convolution(image, self.d_filters[0], name='d_h0_conv')
if not reuse:
self.d_weights.append(weight)
self.d_layers.append(lrelu(conv))
for i in range(0, self.opt.num_layers - 2):
conv, weight = convolution(self.d_layers[-1],
self.d_filters[i+1],
name='d_h'+str(i+1)+'_conv')
if not reuse:
self.d_weights.append(weight)
self.d_layers.append(
lrelu(
self.d_bn_layers[i](conv)
)
)
# last layer
logit, weight = convolution(self.d_layers[-1], self.d_filters[-1], name='d_h4_conv')
if not reuse:
self.d_weights.append(weight)
self.d_layers.append(tf.nn.sigmoid(logit))
return self.d_layers[-1], logit
def generator(self, z, reuse=False, train=True):
"""
In generator network batch normalization used
to all layers except the output layer
:param z:
:param reuse:
:param train:
:return:
"""
with tf.variable_scope("generator") as scope:
if reuse:
scope.reuse_variables()
if train:
_, h, w, in_channels = [i.value for i in z.get_shape()]
else:
sh = tf.shape(z)
h = tf.cast(sh[1], tf.int32)
w = tf.cast(sh[2], tf.int32)
self.g_layers.append(z)
# upscale image num_layers times
for i in range(0, self.opt.num_layers - 1):
new_h = (2**(i+1))*(h-1)+1
new_w = (2**(i+1))*(w-1)+1
out_shape = [self.batch_size, new_h, new_w, self.g_filters[i]]
# deconvolve / upscale 2 times
layer, weight = deconvolution(self.g_layers[-1], out_shape, i, name='g_h'+str(i))
self.g_weights.append(weight)
# batch normalization and activation
self.g_layers.append(tf.nn.relu(self.g_bn_layers[i](layer, train)))
# upscale
layer, weight = deconvolution(self.g_layers[-1],
[
self.batch_size,
(2**self.opt.num_layers)*(h-1)+1,
(2**self.opt.num_layers)*(w-1)+1,
self.g_filters[self.opt.num_layers - 1]
],
self.opt.num_layers - 1,
name='g_h'+str(self.opt.num_layers - 1))
self.g_weights.append(weight)
# activate without batch normalization
self.g_layers.append(tf.nn.tanh(layer, name='output'))
return self.g_layers[-1]
def load(self, checkpoint_dir):
print(" [*] Reading checkpoints...")
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
ckpt_name = os.path.basename(ckpt.model_checkpoint_path)
self.saver.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))
counter = int(next(re.finditer("(\d+)(?!.*\d)",ckpt_name)).group(0))
print(" [*] Success to read {}".format(ckpt_name))
return True, counter
else:
print(" [*] Failed to find a checkpoint")
return False, 0
def hyper_param_summarization(self, h_dict):
# define the list of important hyper_params
for name, var in h_dict:
self.hyper_params += str(name)+'='+str(var)