-
-
Notifications
You must be signed in to change notification settings - Fork 198
/
AnimeGANv3_shinkai.py
373 lines (310 loc) · 19.9 KB
/
AnimeGANv3_shinkai.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
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
from tools.ops import *
from tools.utils import *
from glob import glob
import time
import numpy as np
from joblib import Parallel, delayed
from skimage import segmentation, color
from net import generator
from net.discriminator import D_net
from tools.data_loader import ImageGenerator
from tools.GuidedFilter import guided_filter
from tools.L0_smoothing import L0Smoothing
class AnimeGANv3(object) :
def __init__(self, sess, args):
self.model_name = 'AnimeGANv3'
self.sess = sess
self.checkpoint_dir = args.checkpoint_dir
self.log_dir = args.log_dir
self.dataset_name = args.style_dataset
self.epoch = args.epoch
self.init_G_epoch = args.init_G_epoch
self.batch_size = args.batch_size
self.save_freq = args.save_freq
self.load_or_resume = args.load_or_resume
self.init_G_lr = args.init_G_lr
self.d_lr = args.d_lr
self.g_lr = args.g_lr
self.img_size = args.img_size
self.img_ch = args.img_ch
""" Discriminator """
self.sn = args.sn
self.sample_dir = os.path.join(args.sample_dir, self.model_dir)
check_folder(self.sample_dir)
self.val_real = tf.placeholder(tf.float32, [1, None, None, self.img_ch], name='val_input')
self.real_photo = tf.placeholder(tf.float32, [self.batch_size, self.img_size[0], self.img_size[1], self.img_ch], name='real_photo')
self.photo_superpixel = tf.placeholder(tf.float32, [self.batch_size, self.img_size[0], self.img_size[1], self.img_ch], name='photo_superpixel')
self.fake_superpixel = tf.placeholder(tf.float32, [self.batch_size, self.img_size[0], self.img_size[1], self.img_ch], name='fake_superpixel')
self.fake_NLMean_l0 = tf.placeholder(tf.float32,[self.batch_size, self.img_size[0], self.img_size[1], self.img_ch], name='fake_NLMean_l0')
self.anime = tf.placeholder(tf.float32, [self.batch_size, self.img_size[0], self.img_size[1], self.img_ch], name='anime_image')
self.anime_smooth = tf.placeholder(tf.float32, [self.batch_size, self.img_size[0], self.img_size[1], self.img_ch], name='anime_smooth_image')
self.real_generator = ImageGenerator('./dataset/train_photo', self.img_size, self.batch_size)
self.anime_image_generator = ImageGenerator('./dataset/{}'.format(self.dataset_name + '/style'), self.img_size, self.batch_size)
self.anime_smooth_generator = ImageGenerator('./dataset/{}'.format(self.dataset_name + '/smooth_noise'), self.img_size, self.batch_size)
self.dataset_num = max(self.real_generator.num_images, self.anime_image_generator.num_images)
print()
print("##### Information #####")
print("# dataset : ", self.dataset_name)
print("# max dataset number : ", self.dataset_num)
print("# batch_size : ", self.batch_size)
print("# epoch : ", self.epoch)
print("# init_G_epoch : ", self.init_G_epoch)
print("# training image size [H, W] : ", self.img_size)
print("# init_G_lr,g_lr,d_lr : ", self.init_G_lr,self.g_lr,self.d_lr)
print()
def generator(self, x_init, is_training, reuse=False, scope="generator"):
with tf.variable_scope(scope, reuse=reuse):
fake_s, fake_m = generator.G_net(x_init, is_training)
return fake_s, fake_m
def discriminator(self, x_init, reuse=False, scope="discriminator"):
return D_net(x_init, self.sn, ch=32, reuse=reuse, scope=scope)
##################################################################################
def build_train(self):
""" Define Generator, Discriminator """
self.generated_s, self.generated_m = self.generator(self.real_photo, is_training=True)
self.generated = self.tanh_out_scale(guided_filter(self.sigm_out_scale(self.generated_s),self.sigm_out_scale(self.generated_s), 2, 0.01)) #0.25**2
"""for val"""
self.val_generated_s, self.val_generated_m = self.generator(self.val_real, is_training=False, reuse=True)
self.val_generated = self.tanh_out_scale(guided_filter(self.sigm_out_scale(self.val_generated_s), self.sigm_out_scale(self.val_generated_s), 2, 0.01)) # 0.25**2
# gray maping
self.fake_sty_gray = tf.image.grayscale_to_rgb(tf.image.rgb_to_grayscale(self.generated))
self.anime_sty_gray = tf.image.grayscale_to_rgb(tf.image.rgb_to_grayscale(self.anime))
self.gray_anime_smooth = tf.image.grayscale_to_rgb(tf.image.rgb_to_grayscale(self.anime_smooth))
# support
fake_gray_logit = self.discriminator(self.fake_sty_gray)
anime_gray_logit = self.discriminator(self.anime_sty_gray, reuse=True, )
gray_anime_smooth_logit = self.discriminator(self.gray_anime_smooth, reuse=True, )
# main
generated_m_logit = self.discriminator(self.generated_m, scope="discriminator_main")
fake_NLMean_logit = self.discriminator(self.fake_NLMean_l0, reuse=True, scope="discriminator_main")
""" Define Loss """
# init G
self.Pre_train_G_loss = con_loss(self.real_photo, self.generated) #+ con_loss(self.real_photo, self.generated_m)
# gan
"""support"""
self.con_loss = con_loss(self.real_photo, self.generated, 0.5)
self.s22, self.s33, self.s44 = style_loss_decentralization_3(self.anime_sty_gray, self.fake_sty_gray, [0.1, 2.0, 28])
self.sty_loss = self.s22 + self.s33 + self.s44
self.rs_loss = region_smoothing_loss(self.fake_superpixel, self.generated, 0.8 ) + \
VGG_LOSS(self.photo_superpixel, self.generated) * 0.5
self.color_loss = Lab_color_loss(self.real_photo, self.generated, 8. )
self.tv_loss = 0.0001 * total_variation_loss(self.generated)
self.g_adv_loss = generator_loss(fake_gray_logit)
self.G_support_loss = self.g_adv_loss + self.con_loss + self.sty_loss + self.rs_loss + self.color_loss +self.tv_loss
self.D_support_loss = discriminator_loss(anime_gray_logit, fake_gray_logit) \
+ discriminator_loss_346(gray_anime_smooth_logit) * 5.
"""main"""
self.tv_loss_m = 0.0001 * total_variation_loss(self.generated_m)
self.p4_loss = VGG_LOSS(self.fake_NLMean_l0, self.generated_m) * 0.5
self.p0_loss = L1_loss(self.fake_NLMean_l0, self.generated_m) * 50
self.g_m_loss = generator_loss_m(generated_m_logit) * 0.02
self.G_main_loss = self.g_m_loss + self.p0_loss + self.p4_loss + self.tv_loss_m
self.D_main_loss = discriminator_loss_m(fake_NLMean_logit, generated_m_logit) * 0.1
self.Generator_loss = self.G_support_loss + self.G_main_loss
self.Discriminator_loss = self.D_support_loss + self.D_main_loss
""" Training """
t_vars = tf.trainable_variables()
G_vars = [var for var in t_vars if 'generator' in var.name]
D_vars = [var for var in t_vars if 'discriminator' in var.name]
# init G
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
self.init_G_optim = tf.train.AdamOptimizer(self.init_G_lr, beta1=0.5, beta2=0.999).minimize(self.Pre_train_G_loss, var_list=G_vars)
###
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
self.G_optim = tf.train.AdamOptimizer(self.g_lr , beta1=0.5, beta2=0.999).minimize(self.Generator_loss, var_list=G_vars)
self.D_optim = tf.train.AdamOptimizer(self.d_lr , beta1=0.5, beta2=0.999).minimize(self.Discriminator_loss, var_list=D_vars)
"""" Summary """
#
self.Summary_G_init_loss = tf.summary.scalar("G_init", self.Pre_train_G_loss)
#
self.Summary_G_adv = tf.summary.scalar("G_adv", self.g_adv_loss)
self.Summary_G_con_loss = tf.summary.scalar("con_loss", self.con_loss)
self.Summary_G_rs_loss = tf.summary.scalar("rs_loss", self.rs_loss)
self.Summary_G_sty_loss = tf.summary.scalar("sty_loss", self.sty_loss)
self.Summary_G_color_loss = tf.summary.scalar("color_loss", self.color_loss)
self.Summary_G_tv_loss = tf.summary.scalar("tv_loss", self.tv_loss)
self.Summary_G_loss = tf.summary.scalar("Generator_loss", self.Generator_loss)
self.Summary_D_loss = tf.summary.scalar("Discriminator_loss", self.Discriminator_loss)
#------
self.pretrianed_G_merge = tf.summary.merge([self.Summary_G_init_loss])
self.GD_loss_merge = tf.summary.merge([self.Summary_G_loss,self.Summary_G_adv, self.Summary_G_con_loss, self.Summary_G_rs_loss, self.Summary_G_sty_loss,self.Summary_G_color_loss,self.Summary_G_tv_loss, self.Summary_D_loss])
def train(self):
# initialize all variables
self.sess.run(tf.global_variables_initializer())
# saver to save model
variables = tf.contrib.framework.get_variables_to_restore()
variables_to_resotre = [v for v in variables if 'Adam' not in v.name]
self.saver_load = tf.train.Saver(var_list=variables_to_resotre, max_to_keep=self.epoch)
self.saver = tf.train.Saver(max_to_keep=self.epoch)
# summary writer
# self.writer = tf.summary.FileWriter(self.log_dir + '/' + self.model_dir, self.sess.graph)
""" Input Image"""
real_photo_op, anime_img_op, anime_smooth_op = self.real_generator.load_images(), self.anime_image_generator.load_images(), self.anime_smooth_generator.load_images()
# restore check-point if it exits
could_load, checkpoint_counter = self.load(self.checkpoint_dir)
if could_load:
start_epoch = checkpoint_counter + 1
print(" [*] Load SUCCESS")
else:
start_epoch = 0
print(" [!] Load failed...")
# loop for epoch
steps = int(self.dataset_num / self.batch_size)
for epoch in range(start_epoch, self.epoch):
for idx in range(steps):
start_time = time.time()
real_photo, anime_, anime_smooth_ = self.sess.run([real_photo_op, anime_img_op, anime_smooth_op])
train_feed_dict = {
self.real_photo:real_photo[0],
self.photo_superpixel: real_photo[1],
self.anime:anime_[0],
self.anime_smooth:anime_smooth_[0],
}
""" pre-training G """
if epoch < self.init_G_epoch :
_, init_loss, summary_str = self.sess.run([self.init_G_optim,self.Pre_train_G_loss, self.pretrianed_G_merge], feed_dict = train_feed_dict)
# self.writer.add_summary(summary_str, epoch)
step_time = time.time() - start_time
print("Epoch: %3d, Step: %5d / %5d, time: %.3fs, ETA: %.2fs, Pre_train_G_loss: %.6f" % (
epoch, idx, steps, step_time, step_time*(steps-idx+1), init_loss))
"""style transfer"""
else:
""" Update G """
# output fake image
inter_out_s, inter_out= self.sess.run([self.generated_s, self.generated], feed_dict=train_feed_dict)
# superpixel_batch = self.get_simple_superpixel_improve(inter_out, seg_num=200)
superpixel_batch = self.get_seg(inter_out)
fake_NLMean_batch = self.get_NLMean_l0(inter_out_s)
train_feed_dict.update(
{
self.fake_superpixel: superpixel_batch,
self.fake_NLMean_l0: fake_NLMean_batch,
}
)
_, G_loss, G_support_loss, g_adv_loss, con_loss, rs_loss, sty_loss, s22, s33, s44, color_loss, tv_loss, \
G_main_loss, g_m_loss, p0_loss,p4_loss,tv_loss_m = self.sess.run([self.G_optim,
self.Generator_loss,
self.G_support_loss,
self.g_adv_loss,
self.con_loss,
self.rs_loss,
self.sty_loss, self.s22, self.s33, self.s44,
self.color_loss,
self.tv_loss,
self.G_main_loss,
self.g_m_loss,
self.p0_loss,
self.p4_loss,
self.tv_loss_m
], feed_dict = train_feed_dict)
""" Update D """
_, D_loss, D_support_loss, D_main_loss, summary_str = self.sess.run([self.D_optim,
self.Discriminator_loss,
self.D_support_loss,
self.D_main_loss,
self.GD_loss_merge],
feed_dict=train_feed_dict)
# self.writer.add_summary(summary_str, epoch)
step_time = time.time() - start_time
info = f'Epoch: {epoch:3d}, Step: {idx:5d} /{steps:5d}, time: {step_time:.3f}s, ETA: {step_time*(steps-idx+1):.2f}s, ' + \
f'D_loss:{D_loss:.3f} ~ G_loss: {G_loss:.3f} || ' + \
f'G_support_loss: {G_support_loss:.6f}, g_s_loss: {g_adv_loss:.6f}, con_loss: {con_loss:.6f}, rs_loss: {rs_loss:.6f}, sty_loss: {sty_loss:.6f}, s22: {s22:.6f}, s33: {s33:.6f}, s44: {s44:.6f}, color_loss: {color_loss:.6f}, tv_loss: {tv_loss:.6f} ~ D_support_loss: {D_support_loss:.6f} || ' + \
f'G_main_loss: {G_main_loss:.6f}, g_m_loss: {g_m_loss:.6f}, p0_loss: {p0_loss:.6f}, p4_loss: {p4_loss:.6f}, tv_loss_m: {tv_loss_m:.6f} ~ D_main_loss: {D_main_loss:.6f}'
print(info)
# 2---------------------------------------------------------------------------------
if (epoch + 1) >= self.init_G_epoch and np.mod(epoch + 1, self.save_freq) == 0:
self.save(self.checkpoint_dir, epoch)
if (epoch + 1) >= self.init_G_epoch:
# if (epoch + 1) >= 1:
""" Result Image """
val_files = glob('./dataset/{}/*.*'.format('val'))
save_path = './{}/{:03d}/'.format(self.sample_dir, epoch)
check_folder(save_path)
for i, sample_file in enumerate(val_files):
print('val: '+ str(i) + sample_file)
sample_image = np.asarray(load_test_data(sample_file, self.img_size))
val_real,test_, test_s, test_m = self.sess.run([self.val_real,self.val_generated,self.val_generated_s,self.val_generated_m ],feed_dict = {self.val_real:sample_image} )
save_images(val_real, save_path+'{:03d}_a.jpg'.format(i))
save_images(test_, save_path+'{:03d}_b.jpg'.format(i))
save_images(test_s, save_path+'{:03d}_c.jpg'.format(i))
save_images(test_m, save_path+'{:03d}_d.jpg'.format(i))
@property
def model_dir(self):
return "{}_{}".format(self.model_name, self.dataset_name)
def get_seg(self, batch_image):
def get_superpixel(image):
image = (image + 1.) * 127.5
image = np.clip(image, 0, 255).astype(np.uint8) # [-1. ,1.] ~ [0, 255]
image_seg = segmentation.felzenszwalb(image, scale=5, sigma=0.8, min_size=100)
image = color.label2rgb(image_seg, image, bg_label=-1, kind='avg').astype(np.float32)
image = image / 127.5 - 1.0
return image
num_job = np.shape(batch_image)[0]
batch_out = Parallel(n_jobs=num_job)(delayed(get_superpixel) (image) for image in batch_image)
return np.array(batch_out)
def get_simple_superpixel_improve(self, batch_image, seg_num=200):
def process_slic(image):
seg_label = segmentation.slic(image, n_segments=seg_num, sigma=1, start_label=0,compactness=10, convert2lab=True)
image = color.label2rgb(seg_label, image, bg_label=-1, kind='avg')
return image
num_job = np.shape(batch_image)[0]
batch_out = Parallel(n_jobs=num_job)(delayed(process_slic)(image )for image in batch_image)
return np.array(batch_out)
def get_NLMean_l0(self, batch_image, ):
def process_slic(image):
image = ((image + 1) * 127.5).clip(0, 255).astype(np.uint8)
image = cv2.fastNlMeansDenoisingColored(image, None, 7, 6, 6, 7)
image = L0Smoothing(image/255, 0.005).astype(np.float32) * 2. - 1.
return image.clip(-1., 1.)
num_job = np.shape(batch_image)[0]
batch_out = Parallel(n_jobs=num_job)(delayed(process_slic)(image) for image in batch_image)
return np.array(batch_out)
def save(self, checkpoint_dir, step):
checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
self.saver.save(self.sess, os.path.join(checkpoint_dir, self.model_name + '.model'), global_step=step)
def load(self, checkpoint_dir):
print(" [*] Reading checkpoints...")
checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir)
ckpt = tf.train.get_checkpoint_state(checkpoint_dir) # checkpoint file information
if ckpt and ckpt.model_checkpoint_path:
ckpt_name = os.path.basename(ckpt.model_checkpoint_path) # first line
if "resume" == self.load_or_resume :
self.saver.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))
else:
self.saver_load.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))
counter = int(ckpt_name.split('-')[-1])
print(" [*] Success to read {}".format(os.path.join(checkpoint_dir, ckpt_name)))
return True, counter
else:
print(" [*] Failed to find a checkpoint")
return False, 0
def to_lab(self, x):
"""
@param x: image tensor [-1.0, 1.0]
# @return: image tensor [-1.0, 1.0]
@return: image tensor [0.0, 1.0]
"""
x = (x + 1.0) / 2.0
x = rgb_to_lab(x)
y = tf.concat([tf.expand_dims(x[:, :, :, 0] / 100.,-1), tf.expand_dims((x[:, :, :, 1]+128.)/255.,-1), tf.expand_dims((x[:, :, :, 2]+128.)/255.,-1)], axis=-1)
return y
def sigm_out_scale(self, x):
"""
@param x: image tensor [-1.0, 1.0]
@return: image tensor [0.0, 1.0]
"""
# [-1.0, 1.0] to [0.0, 1.0]
x = (x + 1.0) / 2.0
return tf.clip_by_value(x, 0.0, 1.0)
def tanh_out_scale(self, x):
"""
@param x: image tensor [0.0, 1.0]
@return: image tensor [-1.0, 1.0]
"""
# [0.0, 1.0] to [-1.0, 1.0]
x = (x - 0.5) * 2.0
return tf.clip_by_value(x,-1.0, 1.0)