/
paper_implementation.py
190 lines (167 loc) · 8.51 KB
/
paper_implementation.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
import os
import numpy as np
import tensorflow as tf
import tensorlayer as tl
from matplotlib import pyplot as plt
import skimage
from skimage.transform import resize
from load_data import load_parsed_sod
class ParameterizedUNet:
def __init__(self, patch_size):
# hyperparameters
self.size = patch_size
self.encoder_channels = [64, 64, 128, 128, 256, 256, 512]
self.scene_parsing_decoder_channels = [512, 256, 256, 128, 128, 64, 64]
self.harmonization_decoder_channels = [512, 256, 256, 128, 128, 64, 64]
self.batch_size = None
self.lr = 1e-3
self.translator_hidden = 32
# i/o tensor
# input img should be patches in size 512
self.input_img = tf.placeholder(shape=[None, self.size, self.size, 3], dtype=tf.float32)
# mask: shape and dtype????
self.input_mask = tf.placeholder(shape=[None, self.size, self.size, 1], dtype=tf.float32)
self.truth_img = tf.placeholder(shape=[None, self.size, self.size, 3], dtype=tf.float32)
self.truth_seg = tf.placeholder(shape=[None, self.size, self.size, 1], dtype=tf.float32)
self.output_img = None
self.output_seg = None
self.is_training = True
# internal tensors, set after building
self.down_layers = None
self.up_layers = None
self.l2_loss = None
self.seg_loss = None
self.loss = None
self.optimizer = None
self.train_op = None
self.merged_summary = None
self.global_step = None
self.saver = None
def build(self):
self.down_layers, encoder_out = self._build_encoder(self.input_img, self.input_mask)
with tf.variable_scope('segmentation'):
seg_layers, self.output_seg = self._build_scene_parsing_decoder(encoder_out, self.down_layers, 2)
with tf.variable_scope('harmonization'):
self.output_img = self._build_harmonization_decoder(encoder_out, self.down_layers, seg_layers, 3)
self.l2_loss = tf.losses.mean_squared_error(self.truth_img, self.output_img)
# self.seg_loss = tf.losses.mean_squared_error(self.truth_seg, self.output_seg)
self.seg_loss = 0
self.loss = self.l2_loss + self.seg_loss
tf.summary.scalar('l2-loss', self.l2_loss)
tf.summary.scalar('seg-loss', self.seg_loss)
tf.summary.scalar('loss', self.loss)
self.optimizer = tf.train.AdamOptimizer(self.lr)
# for Batch Norm update
update_ops = tf.compat.v1.get_collection(tf.GraphKeys.UPDATE_OPS)
train_op = self.optimizer.minimize(self.loss,
global_step=tf.train.get_or_create_global_step())
self.train_op = tf.group([train_op, update_ops])
self.merged_summary = tf.summary.merge_all()
self.global_step = tf.train.get_or_create_global_step()
def _build_encoder(self, img, mask):
# down sampling
down_layers = []
tensor = tf.concat([img, mask], axis=3)
tensor = tl.layers.InputLayer(tensor)
for i, c in enumerate(self.encoder_channels):
tensor = tl.layers.Conv2d(tensor, n_filter=c, filter_size=(4, 4), strides=(2, 2), act=tf.nn.elu,
padding='SAME', name="encoder_Conv2d_{}".format(i))
tensor = tl.layers.BatchNormLayer(tensor, is_train=self.is_training, name="encoder_batch_norm_{}".format(i))
down_layers.append(tensor)
# no shrink for the last layer
tensor = tl.layers.FlattenLayer(tensor)
tensor = tl.layers.DenseLayer(tensor, n_units=1024, name="encoder_fc")
tensor = tl.layers.ReshapeLayer(tensor, shape=(-1, 1, 1, 1024), name="encoder_reshape")
return down_layers, tensor
def _build_scene_parsing_decoder(self, encoder_out, down_layers, out_channel):
seg_layers = []
tensor = tl.layers.ReshapeLayer(encoder_out, shape=(-1, 2, 2, 512))
for i, c in enumerate(self.scene_parsing_decoder_channels):
tensor = tl.layers.DeConv2d(tensor, n_filter=c, filter_size=(4, 4), strides=(2, 2),
act=tf.nn.elu, padding='SAME', name="seg_decoder_DeConv2d_{}".format(i))
# print('after', tensor, self.down_layers[-(i + 2)])
tensor = tl.layers.BatchNormLayer(tensor, is_train=self.is_training, name="seg_batch_norm_{}".format(i))
tensor = tl.layers.ElementwiseLayer([tensor, down_layers[-(i + 1)]], combine_fn=tf.add,
name="seg_elesum_{}".format(i))
# tensor = tf.concat([tensor, down_layers[-(i + 2)]], axis=3, name=f'scene_parse_decoder{i}_out')
seg_layers.append(tensor)
scene_parse = tl.layers.Conv2d(tensor, n_filter=out_channel, filter_size=(1, 1), strides=(1, 1),
padding='SAME', name="get_segmentation")
return seg_layers, scene_parse.outputs
def _build_harmonization_decoder(self, encoder_out, down_layers, seg_layers, out_channel):
tensor = tl.layers.ReshapeLayer(encoder_out, shape=(-1, 2, 2, 512))
for i, c in enumerate(self.harmonization_decoder_channels):
tensor = tl.layers.DeConv2d(tensor, n_filter=c, filter_size=(4, 4), strides=(2, 2),
act=tf.nn.elu, padding='SAME', name="harm_DeConv2d_{}".format(i))
# print('after', tensor, self.down_layers[-(i + 2)])
tensor = tl.layers.BatchNormLayer(tensor, is_train=self.is_training, name="harm_batch_norm_{}".format(i))
tensor = tl.layers.ElementwiseLayer([tensor, down_layers[-(i + 1)]], combine_fn=tf.add,
name="harm_elesum_{}".format(i))
tensor = tl.layers.ConcatLayer([tensor, seg_layers[i]], 3, name="concat_{}".format(i))
harm_img = tl.layers.DeConv2d(tensor, n_filter=out_channel, filter_size=(4, 4), strides=(2, 2),
padding='SAME', name='get_img')
return harm_img.outputs
def save(self, sess, path="tmp/paired/", global_step=None):
if not os.path.exists(path):
os.makedirs(path, exist_ok=True)
if self.saver is None:
self.saver = tf.train.Saver(keep_checkpoint_every_n_hours=0.5)
self.saver.save(sess, path, global_step=global_step)
def restore(self, sess, path="tmp/paired/"):
self.saver.restore(sess, path)
def tweak_foreground(image, mask):
"""
tweak foreground by apply random factor
"""
mask = np.expand_dims(mask, 2)
tweaked = mask * image * np.random.uniform(0.1, 2)
new_image = (1 - mask) * image + tweaked
new_image *= (1.0 / new_image.max())
return new_image
def train():
np.random.seed(0)
patch_size = 512
batch_size = 8
images, masks = load_parsed_sod()
images = np.array([resize(im, (patch_size, patch_size)) for im in images])
masks = np.array(
[np.expand_dims(skimage.img_as_bool(resize(skimage.img_as_float(ms), (patch_size, patch_size))), 2) for ms in
masks])
sess = tf.Session()
net = ParameterizedUNet(patch_size=patch_size)
net.build()
sess.run(tf.global_variables_initializer())
print("start training")
for epoch in range(500):
out_im = None
truth_img = None
tweaked = None
truth_mask = None
index = np.arange(len(images))
np.random.shuffle(index)
for i, batch_index in enumerate(np.array_split(index, len(index) // batch_size)):
truth_img = images[batch_index]
truth_mask = masks[batch_index]
tweaked = [tweak_foreground(im, ms) for im, ms in zip(truth_img, truth_mask)]
# TODO: tweaked is in (0, 1) which is good but why
_, loss, out_im = sess.run([net.train_op, net.loss, net.output_img],
feed_dict={net.input_img: tweaked,
net.input_mask: truth_mask,
net.truth_img: truth_img})
print('epoch', epoch, 'batch', i, loss, flush=True)
else:
plt.subplot(2, 3, 1)
plt.imshow(tweaked[0])
plt.title('Input')
plt.subplot(2, 3, 2)
plt.imshow(truth_img[0])
plt.title('Truth')
plt.subplot(2, 3, 3)
plt.imshow(out_im[0])
plt.title('out')
plt.subplot(2, 3, 5)
plt.imshow(truth_mask[0])
plt.savefig(f'tmp/{epoch}.png')
plt.close()
if __name__ == '__main__':
train()