-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathgan.py
186 lines (162 loc) · 8.17 KB
/
gan.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
import tensorflow as tf
import numpy as np
import os
import csv
import time
from PIL import Image
class GAN:
def __init__(self, generator,
discriminator,
metrics='JSD',
lr_d=1e-4,
lr_g=1e-4,
eps=1e-12,
is_training=True):
self.discriminator = discriminator
self.generator = generator
self.image_shape = self.discriminator.input_shape
self.noise_dim = self.generator.noise_dim
self.image = tf.placeholder(tf.float32, [None] + list(self.image_shape), name='x')
self.noise = tf.placeholder(tf.float32, [None, self.noise_dim], name='z')
self.generate = self.generator(self.noise, reuse=False, is_training=True)
self.discriminate_real = self.discriminator(self.image, reuse=False, is_training=True)
self.discriminate_fake = self.discriminator(self.generate, reuse=True, is_training=True)
with tf.name_scope('loss'):
if metrics == 'JSD':
real_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(self.discriminate_real),
logits=self.discriminate_real))
fake_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.zeros_like(self.discriminate_fake),
logits=self.discriminate_fake))
self.loss_d = real_loss + fake_loss
self.loss_g = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(self.discriminate_fake),
logits=self.discriminate_fake))
elif metrics == 'WD':
self.loss_g = -tf.reduce_mean(self.discriminate_fake)
self.loss_d = -(tf.reduce_mean(self.discriminate_real)
- tf.reduce_mean(self.discriminate_fake))
self.loss_d += tf.nn.l2_loss(self.discriminate_real) * 0.1
else:
raise NotImplementedError
# Optimizer
if is_training:
with tf.name_scope('Optimizer'):
if metrics == 'JSD':
self.opt_d = tf.train.AdamOptimizer(learning_rate=lr_d, beta1=0.5, beta2=0.99) \
.minimize(self.loss_d,
var_list=self.discriminator.vars)
self.opt_g = tf.train.AdamOptimizer(learning_rate=lr_g, beta1=0.5, beta2=0.99) \
.minimize(self.loss_g,
var_list=self.generator.vars)
elif metrics == 'WD':
self.opt_d = tf.train.AdamOptimizer(learning_rate=lr_d, beta1=0.5, beta2=0.99) \
.minimize(self.loss_d,
var_list=self.discriminator.vars)
self.opt_g = tf.train.AdamOptimizer(learning_rate=lr_g, beta1=0.5, beta2=0.99) \
.minimize(self.loss_g,
var_list=self.generator.vars)
else:
raise NotImplementedError
self.saver = tf.train.Saver()
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
self.is_training = False
self.logdir = None
self.tb_writer = None
def fit(self, image_sampler,
noise_sampler,
nb_epoch=1000,
visualize_steps=1,
save_steps=1,
logdir='../logs'):
self.logdir = logdir
os.makedirs(logdir, exist_ok=True)
self.tb_writer = tf.summary.FileWriter(logdir, graph=self.sess.graph)
batch_size = image_sampler.batch_size
nb_sample = image_sampler.nb_sample
# prepare for csv
f = open(os.path.join(logdir, 'learning_log.csv'), 'w')
writer = csv.writer(f, lineterminator='\n')
# calc steps_per_epoch
steps_per_epoch = nb_sample // batch_size
if nb_sample % batch_size != 0:
steps_per_epoch += 1
# for display and csv
loss_g = 0
writer.writerow(['loss_d', 'loss_g'])
fixed_noise = noise_sampler(batch_size, self.noise_dim)
# fit loop
for epoch in range(1, nb_epoch + 1):
print('\nepoch {} / {}'.format(epoch, nb_epoch))
start = time.time()
for iter_ in range(1, steps_per_epoch + 1):
image_batch = image_sampler()
# if image_batch.shape[0] != batch_size:
# continue
noise_batch = noise_sampler(image_batch.shape[0], self.noise_dim)
_, loss_d, = self.sess.run([self.opt_d, self.loss_d],
feed_dict={self.image: image_batch,
self.noise: noise_batch})
_, loss_g = self.sess.run([self.opt_g, self.loss_g],
feed_dict={self.noise: noise_batch})
print('iter : {} / {} {:.1f}[s] loss_d : {:.4f} loss_g : {:.4f} \r'
.format(iter_, steps_per_epoch, time.time() - start,
loss_d, loss_g), end='')
writer.writerow([loss_d, loss_g])
if epoch % visualize_steps == 0:
# noise_batch = noise_sampler(batch_size, self.noise_dim)
self.visualize(os.path.join(logdir, 'epoch_{}'.format(epoch)),
fixed_noise, image_sampler.data_to_image)
if epoch % save_steps == 0:
self.save(epoch)
print('\nTraining is done ...\n')
def restore(self, file_path, mode='both'):
assert mode in ['both', 'discriminator']
reader = tf.train.NewCheckpointReader(file_path)
saved_shapes = reader.get_variable_to_shape_map()
var_names = sorted([(var.name, var.name.split(':')[0])
for var in tf.global_variables()
if var.name.split(':')[0] in saved_shapes])
restore_vars = []
var_dict = dict(zip(map(lambda x:
x.name.split(':')[0], tf.global_variables()), tf.global_variables()))
with tf.variable_scope('', reuse=True):
for var_name, saved_var_name in var_names:
if mode == 'discriminator' and 'discriminator' not in var_name:
continue
current_var = var_dict[saved_var_name]
var_shape = current_var.get_shape().as_list()
if var_shape == saved_shapes[saved_var_name]:
restore_vars.append(current_var)
saver = tf.train.Saver(restore_vars)
saver.restore(self.sess, file_path)
def visualize(self, dst_dir, noise_batch, convert_function):
generated_data = self.predict_on_batch(noise_batch)
generated_images = convert_function(generated_data)
if not os.path.exists(dst_dir):
os.makedirs(dst_dir)
for i, image in enumerate(generated_images):
if image.shape[2] == 1:
image = image.reshape(image.shape[:2])
dst_path = os.path.join(dst_dir, "{}.png".format(i))
pil_image = Image.fromarray(np.uint8(image))
pil_image.save(dst_path)
def save(self, epoch):
dst_dir = os.path.join(self.logdir, "epoch_{}".format(epoch))
if not os.path.exists(dst_dir):
os.makedirs(dst_dir)
return self.saver.save(self.sess, save_path=os.path.join(dst_dir, 'model.ckpt'))
def predict(self, x, batch_size=16):
outputs = np.empty([0] + list(self.image_shape))
steps_per_epoch = len(x) // batch_size if len(x) % batch_size == 0 \
else len(x) // batch_size + 1
for iter_ in range(steps_per_epoch):
x_batch = x[iter_ * batch_size: (iter_ + 1) * batch_size]
o = self.predict_on_batch(x_batch)
outputs = np.append(outputs, o, axis=0)
return outputs
def predict_on_batch(self, x):
return self.sess.run(self.generate,
feed_dict={self.noise: x})