-
Notifications
You must be signed in to change notification settings - Fork 15
/
model.py
313 lines (254 loc) · 13.3 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
import tensorflow as tf
from util import *
import random
CROP_SIZE = 224
img_width = 256
img_heigh = 256
# --------------------------------- LALER FUNCTION ----------------------------------------- #
def Conv2d(batch_input, n_fiter, filter_size, strides, act=None, padding='SAME', name='conv'):
with tf.variable_scope(name):
in_channels = batch_input.get_shape()[3]
filters = tf.get_variable('filter', [filter_size, filter_size, in_channels, n_fiter], dtype=tf.float32,
initializer=tf.random_normal_initializer(0, 0.02))
conv = tf.nn.conv2d(batch_input, filters, [1, strides, strides, 1], padding=padding)
if act is not None:
conv = act(conv)
return conv
def Deconv(batch_input, n_fiter, filter_size, strides, act=None, padding='SAME', name='deconv'):
with tf.variable_scope(name):
x_shape = tf.shape(batch_input)
output_shape = tf.stack([x_shape[0], x_shape[1] * 2, x_shape[2] * 2, n_fiter])
in_channels = batch_input.get_shape()[-1]
filters = tf.get_variable('filter', [filter_size, filter_size, n_fiter, in_channels], dtype=tf.float32,
initializer=tf.random_normal_initializer(0, 0.02))
conv = tf.nn.conv2d_transpose(batch_input, filters, output_shape, [1, strides, strides, 1], padding=padding)
conv = tf.reshape(conv, output_shape)
if act is not None:
conv = act(conv)
return conv
def LeakyReLu(x, a):
with tf.name_scope("lrelu"):
# adding these together creates the leak part and linear part
# then cancels them out by subtracting/adding an absolute value term
# leak: a*x/2 - a*abs(x)/2
# linear: x/2 + abs(x)/2
# this block looks like it has 2 inputs on the graph unless we do this
x = tf.identity(x)
return (0.5 * (1 + a)) * x + (0.5 * (1 - a)) * tf.abs(x)
def Batchnorm(input, act, is_train, name):
with tf.variable_scope(name):
# this block looks like it has 3 inputs on the graph unless we do this
input = tf.identity(input)
variance_epsilon = 1e-5
normalized = tf.contrib.layers.batch_norm(input, center=True, scale=True, epsilon=variance_epsilon,
activation_fn=act, is_training=is_train, reuse=None)
'''
channels = input.get_shape()[3]
offset = tf.get_variable("offset", [channels], dtype=tf.float32, initializer=tf.zeros_initializer())
scale = tf.get_variable("scale", [channels], dtype=tf.float32, initializer=tf.random_normal_initializer(1.0, 0.02))
mean, variance = tf.nn.moments(input, axes=[0, 1, 2], keep_dims=False)
variance_epsilon = 1e-5
normalized = tf.nn.batch_normalization(input, mean, variance, offset, scale, variance_epsilon=variance_epsilon)
'''
return normalized
def Elementwise(n1, n2, act, name):
with tf.variable_scope(name):
return act(n1, n2)
# --------------------------------- MODEL DEFINITION --------------------------------- #
def input_producer(data_list, channels, batch_size, need_shuffle):
if len(data_list) == 0:
raise Exception("empty data list!")
#data_list = open(data_list_file, 'rt').read().splitlines()
def read_data(data_queue):
# note : read one training data : pixel range : [0, 255]
in_img = tf.image.decode_image(tf.read_file(data_queue[0]), channels=channels)
gt_img = tf.image.decode_image(tf.read_file(data_queue[1]), channels=channels)
def preprocessing(input):
proc = tf.cast(input, tf.float32)
proc.set_shape([img_width, img_heigh, channels])
# normalization
proc = proc / 127.5 - 1
return proc
# output pixel's range : [-1, 1]
in_imgproc = preprocessing(in_img)
gt_imgproc = preprocessing(gt_img)
return in_imgproc, gt_imgproc
with tf.variable_scope('input'):
# Get full list of image and labels
imglist = [s.split(' ')[0] for s in data_list]
lablist = [s.split(' ')[-1] for s in data_list]
srcfilelist = tf.convert_to_tensor(imglist, dtype=tf.string)
dstfilelist = tf.convert_to_tensor(lablist, dtype=tf.string)
# Put images and label into a queue
data_queue = tf.train.slice_input_producer([srcfilelist, dstfilelist], capacity=64, shuffle=need_shuffle)
# Read one data from queue
input, target = read_data(data_queue)
'''
# testing
with tf.Session() as sess:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
sdfa1, sdfa2 = sess.run([input, target])
print sdfa1.shape, sdfa2.shape
'''
# Construct a batch of training data
in_batch, gt_batch = tf.train.batch([input, target], batch_size, num_threads=1, capacity=64)
return in_batch, gt_batch, len(data_list)
def encode(inputs, out_channels, is_train=False, reuse=False):
with tf.variable_scope("encode", reuse=reuse):
# in
n = Conv2d(inputs, 64, filter_size=3, strides=1, padding='SAME', name='in/k3n64s1')
#n = Batchnorm(n, act=tf.nn.relu, is_train=is_train, name='in/BN')
# start residual blocks
for i in range(2):
nn = Conv2d(n, 64, filter_size=3, strides=1, act=tf.nn.relu, padding='SAME', name='sen64s1/c1/%s' % i)
#nn = Batchnorm(nn, act=tf.nn.relu, is_train=is_train, name='sen64s1/b1/%s' % i)
nn = Conv2d(nn, 64, filter_size=3, strides=1, padding='SAME', name='sen64s1/c2/%s' % i)
#nn = Batchnorm(nn, act=None, is_train=is_train, name='sen64s1/b2/%s' % i)
nn = Elementwise(n, nn, tf.add, 'seb_residual_add/%s' % i)
n = nn
n1 = n
# down size
n = Conv2d(n1, 128, filter_size=3, strides=2, padding='SAME', name='down1-1/k3n128s2')
n2 = Conv2d(n, 128, filter_size=3, strides=1, act=tf.nn.relu, padding='SAME', name='down1-2/k3n128s1')
#n2 = Batchnorm(n, act=tf.nn.relu, is_train=is_train, name='down1/BN')
n = Conv2d(n2, 256, filter_size=3, strides=2, padding='SAME', name='down2-1/k3n256s2')
n = Conv2d(n, 256, filter_size=3, strides=1, act=tf.nn.relu, padding='SAME', name='down2-2/k3n256s1')
#n = Batchnorm(n, act=tf.nn.relu, is_train=is_train, name='down2/BN')
# residual blocks
for i in range(4):
nn = Conv2d(n, 256, filter_size=3, strides=1, act=tf.nn.relu, padding='SAME', name='en64s1/c1/%s' % i)
#nn = Batchnorm(nn, act=tf.nn.relu, is_train=is_train, name='en64s1/b1/%s' % i)
nn = Conv2d(nn, 256, filter_size=3, strides=1, padding='SAME', name='en64s1/c2/%s' % i)
#nn = Batchnorm(nn, act=None, is_train=is_train, name='en64s1/b2/%s' % i)
nn = Elementwise(n, nn, tf.add, 'eb_residual_add/%s' % i)
n = nn
# up size
n = tf.image.resize_images(n, [tf.shape(n)[1] * 2, tf.shape(n)[2] * 2], method=1)
n = Conv2d(n, 128, filter_size=3, strides=1, padding='SAME', name='up1-1/k3n128s1')
n = Conv2d(n, 128, filter_size=3, strides=1, act=tf.nn.relu, padding='SAME', name='up1-2/k3n128s1')
#n = Batchnorm(n, act=tf.nn.relu, is_train=is_train, name='up1/BN')
n = Elementwise(n, n2, tf.add, 'skipping1')
n = tf.image.resize_images(n, [tf.shape(n)[1] * 2, tf.shape(n)[2] * 2], method=1)
n = Conv2d(n, 64, filter_size=3, strides=1, padding='SAME', name='up2-1/k3n64s1')
n = Conv2d(n, 64, filter_size=3, strides=1, act=tf.nn.relu, padding='SAME', name='up2-2/k3n64s1')
#n = Batchnorm(n, act=tf.nn.relu, is_train=is_train, name='up2/BN')
n = Elementwise(n, n1, tf.add, 'skipping2')
# end residual blocks
for i in range(2):
nn = Conv2d(n, 64, filter_size=3, strides=1, act=tf.nn.relu, padding='SAME', name='een64s1/c1/%s' % i)
#nn = Batchnorm(nn, act=tf.nn.relu, is_train=is_train, name='een64s1/b1/%s' % i)
nn = Conv2d(nn, 64, filter_size=3, strides=1, padding='SAME', name='een64s1/c2/%s' % i)
#nn = Batchnorm(nn, act=None, is_train=is_train, name='een64s1/b2/%s' % i)
nn = Elementwise(n, nn, tf.add, 'eeb_residual_add/%s' % i)
n = nn
latent_maps = Conv2d(n, out_channels, filter_size=3, strides=1, act=tf.nn.tanh, padding='SAME', name='latent')
return latent_maps
def decode(latents, out_channels, is_train=False, reuse=False):
with tf.variable_scope("decode", reuse=reuse):
# Decoder
n = Conv2d(latents, 64, filter_size=3, strides=1, padding='SAME', name='in/k3n64s1')
for i in range(8):
nn = Conv2d(n, 64, filter_size=3, strides=1, act=tf.nn.relu, padding='SAME', name='dn64s1/c1/%s' % i)
#nn = Batchnorm(nn, act=tf.nn.relu, is_train=is_train, name='dn64s1/b1/%s' % i)
nn = Conv2d(nn, 64, filter_size=3, strides=1, padding='SAME', name='dn64s1/c2/%s' % i)
#nn = Batchnorm(nn, act=None, is_train=is_train, name='dn64s1/b2/%s' % i)
nn = Elementwise(n, nn, tf.add, 'db_residual_add/%s' % i)
n = nn
n = Conv2d(n, 256, filter_size=3, strides=1, act=None, padding='SAME', name='n256s1/2')
n = Conv2d(n, out_channels, filter_size=1, strides=1, padding='SAME', name='out')
output_map = tf.nn.tanh(n)
return output_map
class VGG19:
def __init__(self, vgg19_npy_path=None):
if vgg19_npy_path is None:
print("Please download vgg19.npz from : https://github.com/machrisaa/tensorflow-vgg")
exit()
self.data_dict = np.load(vgg19_npy_path, encoding='latin1').item()
print("vgg19 npy file loaded!")
def build(self, input, is_rgb):
"""
load variable from npy to build the VGG
:param rgb: rgb image [batch, height, width, 3] values scaled [0, 1]
"""
if is_rgb:
rgb = input
else:
rgb = tf.concat([input, input, input], -1)
shape = rgb.get_shape().as_list()
shape[-1] = 3
rgb.set_shape(shape)
VGG_MEAN = [103.939, 116.779, 123.68]
with tf.name_scope("VGG19"):
rgb_scaled = rgb * 255.0
# Convert RGB to BGR
red, green, blue = tf.split(axis=3, num_or_size_splits=3, value=rgb_scaled)
assert red.get_shape().as_list()[1:] == [224, 224, 1]
assert green.get_shape().as_list()[1:] == [224, 224, 1]
assert blue.get_shape().as_list()[1:] == [224, 224, 1]
bgr = tf.concat(axis=3, values=[
blue - VGG_MEAN[0],
green - VGG_MEAN[1],
red - VGG_MEAN[2],
])
assert bgr.get_shape().as_list()[1:] == [224, 224, 3]
""" conv1 """
nn = self.conv_layer(bgr, "conv1_1")
nn = self.conv_layer(nn, "conv1_2")
nn = self.max_pool(nn, 'pool1')
""" conv2 """
nn = self.conv_layer(nn, "conv2_1")
nn = self.conv_layer(nn, "conv2_2")
nn = self.max_pool(nn, 'pool2')
""" conv3 """
nn = self.conv_layer(nn, "conv3_1")
nn = self.conv_layer(nn, "conv3_2")
nn = self.conv_layer(nn, "conv3_3")
nn = self.conv_layer(nn, "conv3_4")
nn = self.max_pool(nn, 'pool3')
""" conv4 """
nn = self.conv_layer(nn, "conv4_1")
# conv4_1
feature_map = nn
return feature_map
nn = self.conv_layer(nn, "conv4_2")
nn = self.conv_layer(nn, "conv4_3")
nn = self.conv_layer(nn, "conv4_4")
nn = self.max_pool(nn, 'pool4')
""" conv5 """
nn = self.conv_layer(nn, "conv5_1")
nn = self.conv_layer(nn, "conv5_2")
nn = self.conv_layer(nn, "conv5_3")
nn = self.conv_layer(nn, "conv5_4")
nn = self.max_pool(nn, 'pool5')
def avg_pool(self, bottom, name):
return tf.nn.avg_pool(bottom, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name=name)
def max_pool(self, bottom, name):
return tf.nn.max_pool(bottom, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name=name)
def conv_layer(self, bottom, name):
with tf.variable_scope(name):
filt = self.get_conv_filter(name)
conv = tf.nn.conv2d(bottom, filt, [1, 1, 1, 1], padding='SAME')
conv_biases = self.get_bias(name)
bias = tf.nn.bias_add(conv, conv_biases)
relu = tf.nn.relu(bias)
return relu
def fc_layer(self, bottom, name):
with tf.variable_scope(name):
shape = bottom.get_shape().as_list()
dim = 1
for d in shape[1:]:
dim *= d
x = tf.reshape(bottom, [-1, dim])
weights = self.get_fc_weight(name)
biases = self.get_bias(name)
# Fully connected layer. Note that the '+' operation automatically
# broadcasts the biases.
fc = tf.nn.bias_add(tf.matmul(x, weights), biases)
return fc
def get_conv_filter(self, name):
return tf.constant(self.data_dict[name][0], name="filter")
def get_bias(self, name):
return tf.constant(self.data_dict[name][1], name="biases")
def get_fc_weight(self, name):
return tf.constant(self.data_dict[name][0], name="weights")