-
Notifications
You must be signed in to change notification settings - Fork 59
/
model.py
316 lines (258 loc) · 17.2 KB
/
model.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
"""
SRNet - Editing Text in the Wild
The main SRNet model implementation.
Copyright (c) 2019 Netease Youdao Information Technology Co.,Ltd.
Licensed under the GPL License (see LICENSE for details)
Written by Yu Qian
"""
import os
import cv2
import numpy as np
import tensorflow as tf
from loss import build_discriminator_loss, build_generator_loss
import cfg
class SRNet():
def __init__(self, shape = [224, 224], name = ''):
self.name = name
self.cnum = 32
self.graph = tf.Graph()
with self.graph.as_default():
self.i_t = tf.placeholder(dtype = tf.float32, shape = [None] + shape + [3])
self.i_s = tf.placeholder(dtype = tf.float32, shape = [None] + shape + [3])
self.t_sk = tf.placeholder(dtype = tf.float32, shape = [None] + shape + [1])
self.t_t = tf.placeholder(dtype = tf.float32, shape = [None] + shape + [3])
self.t_b = tf.placeholder(dtype = tf.float32, shape = [None] + shape + [3])
self.t_f = tf.placeholder(dtype = tf.float32, shape = [None] + shape + [3])
self.mask_t = tf.placeholder(dtype = tf.float32, shape = [None] + shape + [1])
self.global_step = tf.Variable(tf.constant(0))
self.build_whole_net_with_loss()
self.build_optimizer()
self.build_summary_op()
def _res_block(self, x, activation = tf.nn.leaky_relu, padding = 'SAME', name = 'res_block'):
cnum = x.get_shape().as_list()[-1]
xin = x
x = tf.layers.conv2d(x, cnum // 4, kernel_size = 1, strides = 1, activation = activation, padding = padding, name = name + '_conv1')
x = tf.layers.conv2d(x, cnum // 4, kernel_size = 3, strides = 1, activation = activation, padding = padding, name = name + '_conv2')
x = tf.layers.conv2d(x, cnum, kernel_size = 1, strides = 1, activation = None, padding = padding, name = name + '_conv3')
x = tf.add(xin, x, name = name + '_add')
x = tf.layers.batch_normalization(x, name = name + '_bn')
x = activation(x, name = name + '_out')
return x
def _conv_bn_relu(self, x, cnum = None, activation = tf.nn.leaky_relu, padding = 'SAME', name = 'conv_bn_relu'):
cnum = x.get_shape().as_list()[-1] if cnum is None else cnum
x = tf.layers.conv2d(x, cnum, kernel_size = 3, strides = 1, activation = None, padding = padding, name = name + '_conv')
x = tf.layers.batch_normalization(x, name = name + '_bn')
x = activation(x, name = name + '_out')
return x
def build_res_net(self, x, activation = tf.nn.leaky_relu, padding = 'SAME', name = 'res_net'):
x = self._res_block(x, activation = activation, padding = padding, name = name + '_block1')
x = self._res_block(x, activation = activation, padding = padding, name = name + '_block2')
x = self._res_block(x, activation = activation, padding = padding, name = name + '_block3')
x = self._res_block(x, activation = activation, padding = padding, name = name + '_block4')
return x
def build_encoder_net(self, x, activation = tf.nn.leaky_relu, padding = 'SAME', name = 'encoder_net', get_feature_map = False):
x = self._conv_bn_relu(x, self.cnum, name = name + '_conv1_1')
x = self._conv_bn_relu(x, self.cnum, name = name + '_conv1_2')
x = tf.layers.conv2d(x, 2 * self.cnum, kernel_size = 3, strides = 2, activation = activation, padding = padding, name = name + '_pool1')
x = self._conv_bn_relu(x, 2 * self.cnum, name = name + '_conv2_1')
x = self._conv_bn_relu(x, 2 * self.cnum, name = name + '_conv2_2')
f1 = x
x = tf.layers.conv2d(x, 4 * self.cnum, kernel_size = 3, strides = 2, activation = activation, padding = padding, name = name + '_pool2')
x = self._conv_bn_relu(x, 4 * self.cnum, name = name + '_conv3_1')
x = self._conv_bn_relu(x, 4 * self.cnum, name = name + '_conv3_2')
f2 = x
x = tf.layers.conv2d(x, 8 * self.cnum, kernel_size = 3, strides = 2, activation = activation, padding = padding, name = name + '_pool3')
x = self._conv_bn_relu(x, 8 * self.cnum, name = name + '_conv4_1')
x = self._conv_bn_relu(x, 8 * self.cnum, name = name + '_conv4_2')
if get_feature_map:
return x, [f2, f1]
else:
return x
def build_decoder_net(self, x, fuse = None, activation = tf.nn.leaky_relu, padding = 'SAME', name = 'decoder_net', get_feature_map = False):
if fuse and fuse[0] is not None:
x = tf.concat([x, fuse[0]], axis = -1, name = name + '_fuse1')
x = self._conv_bn_relu(x, 8 * self.cnum, name = name + '_conv1_1')
x = self._conv_bn_relu(x, 8 * self.cnum, name = name + '_conv1_2')
f1 = x
x = tf.layers.conv2d_transpose(x, 4 * self.cnum, kernel_size = 3, strides = 2, activation = activation, padding = padding, name = name + '_deconv1')
if fuse and fuse[1] is not None:
x = tf.concat([x, fuse[1]], axis = -1, name = name + '_fuse2')
x = self._conv_bn_relu(x, 4 * self.cnum, name = name + '_conv2_1')
x = self._conv_bn_relu(x, 4 * self.cnum, name = name + '_conv2_2')
f2 = x
x = tf.layers.conv2d_transpose(x, 2 * self.cnum, kernel_size = 3, strides = 2, activation = activation, padding = padding, name = name + '_deconv2')
if fuse and fuse[2] is not None:
x = tf.concat([x, fuse[2]], axis = -1, name = name + '_fuse3')
x = self._conv_bn_relu(x, 2 * self.cnum, name = name + '_conv3_1')
x = self._conv_bn_relu(x, 2 * self.cnum, name = name + '_conv3_2')
f3 = x
x = tf.layers.conv2d_transpose(x, self.cnum, kernel_size = 3, strides = 2, activation = activation, padding = padding, name = name + '_deconv3')
x = self._conv_bn_relu(x, self.cnum, name = name + '_conv4_1')
x = self._conv_bn_relu(x, self.cnum, name = name + '_conv4_2')
if get_feature_map:
return x, [f1, f2, f3]
else:
return x
def build_text_conversion_net(self, x_t, x_s, padding = 'SAME', name = 'tcn'):
x_t = self.build_encoder_net(x_t, name = name + '_t_encoder')
x_t = self.build_res_net(x_t, name = name + '_t_res')
x_s = self.build_encoder_net(x_s, name = name + '_s_encoder')
x_s = self.build_res_net(x_s, name = name + '_s_res')
x = tf.concat([x_t, x_s], axis = -1, name = name + '_concat1')
y_sk = self.build_decoder_net(x, name = name + '_sk_decoder')
y_sk_out = tf.layers.conv2d(y_sk, 1, kernel_size = 3, strides = 1, activation = 'sigmoid', padding = padding, name = name + '_sk_out')
y_t = self.build_decoder_net(x, name = name + '_t_decoder')
y_t = tf.concat([y_sk, y_t], axis = -1, name = name + '_concat2')
y_t = self._conv_bn_relu(y_t, name = name + '_t_cbr')
y_t_out = tf.layers.conv2d(y_t, 3, kernel_size = 3, strides = 1, activation = 'tanh', padding = padding, name = name + '_t_out')
return y_sk_out, y_t_out
def build_background_inpainting_net(self, x, padding = 'SAME', name = 'bin'):
x, f_encoder = self.build_encoder_net(x, name = name + '_encoder', get_feature_map = True)
x = self.build_res_net(x, name = name + '_res')
x, fuse = self.build_decoder_net(x, fuse = [None] + f_encoder, name = name + '_decoder', get_feature_map = True)
x = tf.layers.conv2d(x, 3, kernel_size = 3, strides = 1, activation = 'tanh', padding = padding, name = name + '_out')
return x, fuse
def build_fusion_net(self, x, fuse, padding = 'SAME', name = 'fn'):
x = self.build_encoder_net(x, name = name + '_encoder')
x = self.build_res_net(x, name = name + '_res')
x = self.build_decoder_net(x, fuse, name = name + '_decoder')
x = tf.layers.conv2d(x, 3, kernel_size = 3, strides = 1, activation = 'tanh', padding = padding, name = name + '_out')
return x
def build_discriminator(self, x, activation = tf.nn.leaky_relu, padding = 'SAME', name = 'discriminator'):
with tf.variable_scope('D'):
x = tf.layers.conv2d(x, 64, kernel_size = 3, strides = 2, activation = activation, padding = padding, name = name + '_conv1')
x = tf.layers.conv2d(x, 128, kernel_size = 3, strides = 2, activation = None, padding = padding, name = name + '_conv2')
x = tf.layers.batch_normalization(x, name = name + '_conv2_bn')
x = activation(x, name = name + '_conv2_activation')
x = tf.layers.conv2d(x, 256, kernel_size = 3, strides = 2, activation = None, padding = padding, name = name + '_conv3')
x = tf.layers.batch_normalization(x, name = name + '_conv3_bn')
x = activation(x, name = name + '_conv3_activation')
x = tf.layers.conv2d(x, 512, kernel_size = 3, strides = 2, activation = None, padding = padding, name = name + '_conv4')
x = tf.layers.batch_normalization(x, name = name + '_conv4_bn')
x = activation(x, name = name + '_conv4_activation')
x = tf.layers.conv2d(x, 1, kernel_size = 3, strides = 1, activation = None, padding = padding, name = name + '_conv5')
x = tf.layers.batch_normalization(x, name = name + '_conv5_bn')
x = tf.nn.sigmoid(x, name = '_out')
return x
def build_generator(self, inputs, name = 'generator'):
i_t, i_s = inputs
with tf.variable_scope('G'):
o_sk, o_t = self.build_text_conversion_net(i_t, i_s, name = name + '_tcn')
o_b, fuse = self.build_background_inpainting_net(i_s, name = name + '_bin')
o_f = self.build_fusion_net(o_t, fuse, name = name + '_fn')
return o_sk, o_t, o_b, o_f
def build_whole_net_with_loss(self):
i_t, i_s = self.i_t, self.i_s
t_sk, t_t, t_b, t_f, mask_t = self.t_sk, self.t_t, self.t_b, self.t_f, self.mask_t
inputs = [i_t, i_s]
labels = [t_sk, t_t, t_b, t_f]
o_sk, o_t, o_b, o_f = self.build_generator(inputs)
self.o_sk = tf.identity(o_sk, name = 'o_sk')
self.o_t = tf.identity(o_t, name = 'o_t')
self.o_b = tf.identity(o_b, name = 'o_b')
self.o_f = tf.identity(o_f, name = 'o_f')
i_db_true = tf.concat([t_b, i_s], axis = -1, name = 'db_true_concat')
i_db_pred = tf.concat([o_b, i_s], axis = -1, name = 'db_pred_concat')
i_db = tf.concat([i_db_true, i_db_pred], axis = 0, name = 'db_concat')
i_df_true = tf.concat([t_f, i_t], axis = -1, name = 'df_true_concat')
i_df_pred = tf.concat([o_f, i_t], axis = -1, name = 'df_pred_concat')
i_df = tf.concat([i_df_true, i_df_pred], axis = 0, name = 'df_concat')
o_db = self.build_discriminator(i_db, name = 'db')
o_df = self.build_discriminator(i_df, name = 'df')
i_vgg = tf.concat([t_f, o_f], axis = 0, name = 'vgg_concat')
vgg_graph_def = tf.GraphDef()
vgg_graph_path = cfg.vgg19_weights
with open(vgg_graph_path, 'rb') as f:
vgg_graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(vgg_graph_def, input_map = {"inputs:0": i_vgg})
with tf.Session() as sess:
o_vgg_1 = sess.graph.get_tensor_by_name("import/block1_conv1/Relu:0")
o_vgg_2 = sess.graph.get_tensor_by_name("import/block2_conv1/Relu:0")
o_vgg_3 = sess.graph.get_tensor_by_name("import/block3_conv1/Relu:0")
o_vgg_4 = sess.graph.get_tensor_by_name("import/block4_conv1/Relu:0")
o_vgg_5 = sess.graph.get_tensor_by_name("import/block5_conv1/Relu:0")
out_g = [o_sk, o_t, o_b, o_f, mask_t]
out_d = [o_db, o_df]
out_vgg = [o_vgg_1, o_vgg_2, o_vgg_3, o_vgg_4, o_vgg_5]
db_loss = build_discriminator_loss(o_db, name = 'db_loss')
df_loss = build_discriminator_loss(o_df, name = 'df_loss')
self.d_loss_detail = [db_loss, df_loss]
self.d_loss = tf.add(db_loss, df_loss, name = 'd_loss')
self.g_loss, self.g_loss_detail = build_generator_loss(out_g, out_d, out_vgg, labels, name = 'g_loss')
def build_optimizer(self):
self.learning_rate = tf.train.exponential_decay(learning_rate = cfg.learning_rate, global_step = self.global_step,
decay_steps = cfg.decay_steps, decay_rate = cfg.decay_rate, staircase = cfg.staircase)
d_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope='D')
with tf.control_dependencies(d_update_ops):
self.d_train_step = tf.train.AdamOptimizer(self.learning_rate, cfg.beta1, cfg.beta2).minimize(self.d_loss,
var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope = 'D'))
g_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope='G')
with tf.control_dependencies(g_update_ops):
self.g_train_step = tf.train.AdamOptimizer(self.learning_rate, cfg.beta1, cfg.beta2).minimize(self.g_loss,
var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope = 'G'))
def build_summary_op(self):
d_summary_loss = tf.summary.scalar("loss", self.d_loss)
d_summary_loss_db = tf.summary.scalar("l_db", self.d_loss_detail[0])
d_summary_loss_df = tf.summary.scalar("l_df", self.d_loss_detail[1])
g_summary_loss = tf.summary.scalar("loss", self.g_loss)
g_summary_loss_t_sk = tf.summary.scalar("l_t_sk", self.g_loss_detail[0])
g_summary_loss_t_l1 = tf.summary.scalar("l_t_l1", self.g_loss_detail[1])
g_summary_loss_b_gan = tf.summary.scalar("l_b_gan", self.g_loss_detail[2])
g_summary_loss_b_l1 = tf.summary.scalar("l_b_l1", self.g_loss_detail[3])
g_summary_loss_f_gan = tf.summary.scalar("l_f_gan", self.g_loss_detail[4])
g_summary_loss_f_l1 = tf.summary.scalar("l_f_l1", self.g_loss_detail[5])
g_summary_loss_f_vgg_per = tf.summary.scalar("l_f_vgg_per", self.g_loss_detail[6])
g_summary_loss_f_vgg_style = tf.summary.scalar("l_f_vgg_style", self.g_loss_detail[7])
self.d_summary_op = tf.summary.merge([d_summary_loss, d_summary_loss_db, d_summary_loss_df])
self.g_summary_op = tf.summary.merge([g_summary_loss, g_summary_loss_t_sk, g_summary_loss_t_l1,
g_summary_loss_b_gan, g_summary_loss_b_l1, g_summary_loss_f_gan,
g_summary_loss_f_l1, g_summary_loss_f_vgg_per, g_summary_loss_f_vgg_style])
self.d_writer = tf.summary.FileWriter(os.path.join(cfg.tensorboard_dir, self.name, 'descriminator'), self.graph)
self.g_writer = tf.summary.FileWriter(os.path.join(cfg.tensorboard_dir, self.name, 'generator'), self.graph)
def train_step(self, sess, global_step, i_t, i_s, t_sk, t_t, t_b, t_f, mask_t):
feed_dict = {
self.i_t: i_t,
self.i_s: i_s,
self.t_sk: t_sk,
self.t_t: t_t,
self.t_b: t_b,
self.t_f: t_f,
self.mask_t: mask_t,
self.global_step: global_step
}
with self.graph.as_default():
_, d_loss, d_log = sess.run([self.d_train_step, self.d_loss, self.d_summary_op], feed_dict = feed_dict)
_, g_loss, g_log = sess.run([self.g_train_step, self.g_loss, self.g_summary_op], feed_dict = feed_dict)
return d_loss, g_loss, d_log, g_log
def predict(self, sess, i_t, i_s, to_shape = None):
assert i_t.shape == i_s.shape and i_t.dtype == i_s.dtype
assert len(i_t.shape) == 3 or (len(i_t.shape) == 4 and to_shape is not None \
and i_t.shape[1] == cfg.data_shape[0] \
and i_t.shape[2] % 8 == 0 \
and i_t.dtype == np.float32)
assert i_t.dtype == np.uint8 \
or (i_t.dtype == np.float32 and np.min(i_t) >= -1 and np.max(i_t) <= 1)
# process raw image, len(i_t.shape) == 3
if len(i_t.shape) == 3:
if not to_shape:
h, w = i_t.shape[:2]
to_shape = (w, h) # w first for cv2
if i_t.shape[0] != cfg.data_shape[0]:
ratio = cfg.data_shape[0] / h
predict_h = cfg.data_shape[0]
predict_w = round(int(w * ratio) / 8) * 8
predict_scale = (predict_w, predict_h) # w first for cv2
i_t = cv2.resize(i_t, predict_scale)
i_s = cv2.resize(i_s, predict_scale)
if i_t.dtype == np.uint8:
i_t = i_t.astype(np.float32) / 127.5 - 1.
i_s = i_s.astype(np.float32) / 127.5 - 1.
i_t = np.expand_dims(i_t, axis = 0)
i_s = np.expand_dims(i_s, axis = 0)
result = sess.run([self.o_sk, self.o_t, self.o_b, self.o_f], feed_dict = {self.i_t: i_t, self.i_s: i_s})
o_sk, o_t, o_b, o_f = result
o_sk = cv2.resize((o_sk[0] * 255.).astype(np.uint8), to_shape, interpolation=cv2.INTER_NEAREST)
o_t = cv2.resize(((o_t[0] + 1.) * 127.5).astype(np.uint8), to_shape)
o_b = cv2.resize(((o_b[0] + 1.) * 127.5).astype(np.uint8), to_shape)
o_f = cv2.resize(((o_f[0] + 1.) * 127.5).astype(np.uint8), to_shape)
return [o_sk, o_t, o_b, o_f]