-
Notifications
You must be signed in to change notification settings - Fork 36
/
OMCNN_2CLSTM.py
254 lines (226 loc) · 14.7 KB
/
OMCNN_2CLSTM.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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
import numpy as np
import BasicNet
import BasicConvLSTMCell
class Net(BasicNet.BasicNet):
# image_size = 448
batch_size = 16
framenum = 16
maximgbatch = 4
init_learning_rate = 10**(-4)
eps = 1e-7
gapnum = 5
salmask_lb = 0.5 #mask cam be salmask_lb~1
dp_in = 0.25
dp_h = 0.25
# num_classes = 20
cell_size = 7
# boxes_per_cell = 2
def __init__(self):
super(Net, self).__init__() # init the fatther class of YoloTinyNet
self.global_step = tf.Variable(0, trainable=False)
self.initial_var_collection.append(self.global_step )
self.out = []
self.predict = []
self.loss = []
self.loss_gt = []
self.re = []
self.loss_gt2 = []
self.yolofeatures_colllection = []
self.flowfeatures_colllection = []
self.startflagcnn = True
#process params
def YOLO_tiny_inference(self, images): # pre128
cnnpretrain = True
cnntrainable = False
self.batch_size = images.get_shape()[0].value
conv_1 = self.conv_layer('conv1', images, 3, 16, stride=1, pretrain=cnnpretrain, batchnormalization=True,
trainable=cnntrainable)
pool_2 = self.max_pool('pool2', conv_1, 2, stride=2)
conv_3 = self.conv_layer('conv3', pool_2, 3, 32, stride=1, pretrain=cnnpretrain, batchnormalization=True,
trainable=cnntrainable)
pool_4 = self.max_pool('pool4', conv_3, 2, stride=2)
conv_5 = self.conv_layer('conv5', pool_4, 3, 64, stride=1, pretrain=cnnpretrain, batchnormalization=True,
trainable=cnntrainable)
pool_6 = self.max_pool('pool6', conv_5, 2, stride=2)
conv_7 = self.conv_layer('conv7', pool_6, 3, 128, stride=1, pretrain=cnnpretrain, batchnormalization=True,
trainable=cnntrainable)
pool_8 = self.max_pool('pool8', conv_7, 2, stride=2)
conv_9 = self.conv_layer('conv9', pool_8, 3, 256, stride=1, pretrain=cnnpretrain, batchnormalization=True,
trainable=cnntrainable)
pool_10 = self.max_pool('pool10', conv_9, 2, stride=2)
conv_11 = self.conv_layer('conv11', pool_10, 3, 512, stride=1, pretrain=cnnpretrain, batchnormalization=True,
trainable=cnntrainable)
pool_12 = self.max_pool('pool12', conv_11, 2, stride=2)
conv_13 = self.conv_layer('conv13', pool_12, 3, 1024, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
conv_14 = self.conv_layer('conv14', conv_13, 3, 1024, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
conv_15 = self.conv_layer('conv15', conv_14, 3, 1024, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
temp_conv = tf.transpose(conv_15, (0, 3, 1, 2))
fc_16 = self.fc_layer('fc16', temp_conv, 256, flat=True, pretrain=cnnpretrain, trainable=cnntrainable)
fc_17 = self.fc_layer('fc17', fc_16, 4096, flat=False, pretrain=cnnpretrain, trainable=cnntrainable)
fc_18 = self.fc_layer('fc18', fc_17, 1470, flat=False, linear=True, pretrain=cnnpretrain, trainable=cnntrainable)
highFeature = tf.reshape(fc_18, [fc_18.get_shape()[0].value, self.cell_size, self.cell_size, -1])
conv_15_2 = self.conv_layer('conv_15_2', conv_15, 1, 128, stride=1,pretrain=cnnpretrain, trainable=cnntrainable)
conv_11_2 = self.conv_layer('conv_11_2', conv_11, 1, 128, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
conv_9_2 = self.conv_layer('conv_9_2', conv_9, 1, 128, stride=1,pretrain=cnnpretrain, trainable=cnntrainable)
# conv_7_2 = self.conv_layer('conv_7_2', conv_7, 1, 256, stride=1, pretrain=False)
tempsize = conv_9.get_shape().as_list()
newconv_7 = tf.image.resize_images(conv_7, [tempsize[1], tempsize[2]])
newconv_9 = tf.image.resize_images(conv_9_2, [tempsize[1], tempsize[2]])
newconv_11_2 = tf.image.resize_images(conv_11_2, [tempsize[1], tempsize[2]])
newconv_15_2 = tf.image.resize_images(conv_15_2, [tempsize[1], tempsize[2]])
highFeature = tf.image.resize_images(highFeature, [tempsize[1], tempsize[2]])
FeatureMap = tf.concat([newconv_7, newconv_9, newconv_11_2, newconv_15_2, highFeature], axis=3)
weight_mask = tf.constant(self.get_centermask(FeatureMap.get_shape().as_list()), dtype=FeatureMap.dtype)
FeatureMap = FeatureMap * weight_mask
return FeatureMap
def Coarse_salmap(self, Yolofeature): # tiny pregen256 fea28_128
cnnpretrain = True
cnntrainable = False
conv_19 = self.conv_layer('conv_19', Yolofeature, 3, 512, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
conv_20 = self.conv_layer('conv_20', conv_19, 1, 256, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
conv_21 = self.conv_layer('conv_21', conv_20, 3, 128, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
conv_22 = self.conv_layer('conv_22', conv_21, 1, 128, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
deconv_23 = self.transpose_conv_layer('deconv_23', conv_22, 4, 16, stride=2, pretrain=cnnpretrain, trainable=cnntrainable)
deconv_24 = self.transpose_conv_layer('deconv_24', deconv_23, 4, 1, stride=2, linear=True, pretrain=cnnpretrain, trainable=cnntrainable)
return deconv_24
def Final_inference(self, cat1, cat2):
cnnpretrain = True
cnntrainable = False
MyFeature = tf.concat([cat1, cat2], axis=3)
Lastconv_1 = self.conv_layer('Lastconv_1', MyFeature, 3, 512, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
Lastconv_2 = self.conv_layer('Lastconv_2', Lastconv_1, 1, 512, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
Lastconv_3 = self.conv_layer('Lastconv_3', Lastconv_2, 3, 256, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
Lastconv_4 = self.conv_layer('Lastconv_4', Lastconv_3, 1, 128, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
return Lastconv_4
def flownet_with_conv(self, x1, x2, mask):
cnnpretrain = True
cnntrainable = False
input = tf.concat([x1, x2], axis=3, name='FNinput')
conv_1 = self.leaky_conv(input, 64, 7, 2, 'FNconv1', pretrain=cnnpretrain, trainable=cnntrainable)
conv_1 = self.conv_mask(conv_1, mask)
conv_2 = self.leaky_conv(conv_1, 128, 5, 2, 'FNconv2', pretrain=cnnpretrain, trainable=cnntrainable)
conv_2 = self.conv_mask(conv_2, mask)
conv_3 = self.leaky_conv(conv_2, 256, 5, 2, 'FNconv3', pretrain=cnnpretrain, trainable=cnntrainable)
conv_3 = self.conv_mask(conv_3, mask)
conv_3_1 = self.leaky_conv(conv_3, 256, 3, 1, 'FNconv3_1', pretrain=cnnpretrain, trainable=cnntrainable)
conv_3_1 = self.conv_mask(conv_3_1, mask)
conv_4 = self.leaky_conv(conv_3_1, 512, 3, 2, 'FNconv4', pretrain=cnnpretrain, trainable=cnntrainable)
conv_4 = self.conv_mask(conv_4, mask)
conv_4_1 = self.leaky_conv(conv_4, 512, 3, 1, 'FNconv4_1', pretrain=cnnpretrain, trainable=cnntrainable)
# conv_4_1 = self.conv_mask(conv_4_1, mask)
conv_5 = self.leaky_conv(conv_4_1, 512, 3, 2, 'FNconv5', pretrain=cnnpretrain, trainable=cnntrainable)
# conv_5 = self.conv_mask(conv_5, mask)
conv_5_1 = self.leaky_conv(conv_5, 512, 3, 1, 'FNconv5_1', pretrain=cnnpretrain, trainable=cnntrainable)
# conv_5_1 = self.conv_mask(conv_5_1, mask)
conv_6 = self.leaky_conv(conv_5_1, 1024, 3, 2, 'FNconv6', pretrain=cnnpretrain, trainable=cnntrainable)
# conv_6 = self.conv_mask(conv_6, mask)
conv_6_1 = self.leaky_conv(conv_6, 1024, 3, 1, 'FNconv6_1', pretrain=cnnpretrain, trainable=cnntrainable)
# conv_6_1 = self.conv_mask(conv_6_1, mask)
out_cat_size = conv_4.get_shape().as_list()
Downconv_6_1 = self.conv_layer('FNDownconv_6_1', conv_6_1, 3, 128, stride=1,pretrain=cnnpretrain, trainable=cnntrainable)
Downconv_5_1 = self.conv_layer('FNDownconv_5_1', conv_5_1, 3, 128, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
Downconv_4_1 = self.conv_layer('FNDownconv_4_1', conv_4_1, 3, 128, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
Downconv_3_1 = self.conv_layer('FNDownconv_3_1', conv_3_1, 3, 128, stride=1, pretrain=cnnpretrain, trainable=cnntrainable)
conv_6_1_cat = tf.image.resize_images(Downconv_6_1, [out_cat_size[1],out_cat_size[2]])
conv_5_1_cat = tf.image.resize_images(Downconv_5_1, [out_cat_size[1],out_cat_size[2]])
conv_4_1_cat = tf.image.resize_images(Downconv_4_1, [out_cat_size[1],out_cat_size[2]])
conv_3_1_cat = tf.image.resize_images(Downconv_3_1, [out_cat_size[1],out_cat_size[2]])
concat_out = tf.concat([conv_6_1_cat, conv_5_1_cat, conv_4_1_cat, conv_3_1_cat], axis=3, name='FNconcat_out')
return concat_out
def inference(self, videoslides, mask_in, mask_h): #videoslides: [batch framenum h w num_features]
with tf.variable_scope('inference'):
shapes = videoslides.get_shape().as_list()
#shapes2 = GTs.get_shape().as_list()
assert len(shapes)==5
self.batch_size = videoslides.get_shape()[0].value
#self.framenum = videoslides.get_shape()[1].value
# assert self.framenum % self.maximgbatch == 0 # frmaenum shoube be the multiple of maximgbatch scope = 'layer_1'
with tf.variable_scope('conv_lstm', initializer=tf.random_uniform_initializer(-.01, 0.1)):
# cell_1 = BasicConvLSTMCell.BasicConvLSTMCell([56, 56], [3, 3], 128, state_is_tuple = False) # input size,fliter size, input channals
# cell_2 = BasicConvLSTMCell.BasicConvLSTMCell([56, 56], [3, 3], 128, state_is_tuple = False) # input size,fliter size, input channals
cell_1 = BasicConvLSTMCell.BasicConvLSTMCell([28, 28], [3, 3], 128,
state_is_tuple=False) # input size,fliter size, input channals
cell_2 = BasicConvLSTMCell.BasicConvLSTMCell([28, 28], [3, 3], 128,
state_is_tuple=False) # input size,fliter size, input channals
new_state_1 = cell_1.zero_state(self.batch_size, 2, tf.float32)
new_state_2 = cell_2.zero_state(self.batch_size, 2, tf.float32)
# print(videoslides.get_shape().as_list())
for indexframe in range(self.framenum):
frame = videoslides[:, indexframe, ...]
#print(indexframe+self.gapnum)
frame_gap = videoslides[:, indexframe+self.gapnum, ...]
#GTframe = GTs[:, indexframe, ...]
Yolo_features = self.YOLO_tiny_inference(frame)
Presalmap = self.Coarse_salmap(Yolo_features)
if self.startflagcnn == True:
self.yolofeatures_colllection = self.pretrain_var_collection
self.pretrain_var_collection = []
salmask = self._normlized_0to1(Presalmap)
salmask = salmask*(1-self.salmask_lb)+self.salmask_lb
Flow_features = self.flownet_with_conv(frame, frame_gap, salmask)
CNNout = self.Final_inference(Yolo_features, Flow_features)
if self.startflagcnn == True:
self.flowfeatures_colllection = self.pretrain_var_collection
y_1, new_state_1 = cell_1(CNNout, new_state_1,mask_in[...,0:4], mask_h[...,0:4], self.dp_in, self.dp_h, 'lstm_layer1')
y_2, new_state_2 = cell_2(y_1, new_state_2,mask_in[...,4:8], mask_h[...,4:8], self.dp_in, self.dp_h, 'lstm_layer2')
deconv = self.transpose_conv_layer('deconv', y_2, 4, 16, stride=2, pretrain=False, trainable=True)
deconv2 = self.transpose_conv_layer('deconv2', deconv, 4, 1, stride=2, linear=True, pretrain=False, trainable=True)
if self.startflagcnn == True:
tf.get_variable_scope().reuse_variables()
self.trainable_var_collection.extend(cell_1.trainable_var_collection)
self.trainable_var_collection.extend(cell_2.trainable_var_collection)
self.startflagcnn = False
output = self._normlized_0to1(deconv2)
#norm_GT = self._normlized(GTframe)
norm_output = self._normlized(output)
# frame_loss = norm_GT * tf.log(self.eps + norm_GT / (norm_output + self.eps))
#frame_loss = tf.reduce_sum(frame_loss) / norm_GT.get_shape()[0].value
# tf.add_to_collection('losses', frame_loss)
output = tf.expand_dims(output, 1)
if indexframe == 0:
tempout = output
else:
tempout = tf.concat([tempout, output], axis=1)
self.out = tempout
def _normlized(self, mat): # tensor [batch_size, image_height, image_width, channels] normalize each fea map
mat_shape = mat.get_shape().as_list()
tempsum = tf.reduce_sum(mat, axis=1)
tempsum = tf.reduce_sum(tempsum, axis=1) + self.eps
tempsum = tf.reshape(tempsum, [-1, 1, 1, mat_shape[3]])
return mat / tempsum
def _normlized_0to1(self, mat): # tensor [batch_size, image_height, image_width, channels] normalize each fea map
mat_shape = mat.get_shape().as_list()
tempmin = tf.reduce_min(mat, axis=1)
tempmin= tf.reduce_min(tempmin, axis=1)
tempmin = tf.reshape(tempmin, [-1, 1, 1, mat_shape[3]])
tempmat = mat - tempmin
tempmax = tf.reduce_max(tempmat, axis=1)
tempmax = tf.reduce_max(tempmax, axis=1) + self.eps
tempmax = tf.reshape(tempmax, [-1, 1, 1, mat_shape[3]])
return tempmat / tempmax
def _loss(self):
weight_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES, scope=None)
loss_weight = tf.add_n(weight_loss)
loss_kl = tf.get_collection('losses', scope=None)
loss_kl = tf.add_n(loss_kl)/self.framenum
# self.out = self.predict
tf.summary.scalar('loss_weight', loss_weight)
tf.summary.scalar('loss_kl', loss_kl)
self.loss_gt = loss_kl
self.loss = loss_kl + loss_weight
def _train(self):
# learning_rate = tf.train.exponential_decay(self.init_learning_rate, self.global_step,
# 100000, 0.95, staircase=True)
# with tf.variable_scope('trainer'):
opt = tf.train.AdamOptimizer(self.init_learning_rate,beta1=0.9, beta2=0.999, epsilon=1e-08)
# for var in self.trainable_var_collection:
# print(var.op.name)
grads = opt.compute_gradients(self.loss,var_list = self.trainable_var_collection)
apply_gradient_op = opt.apply_gradients(grads, global_step=self.global_step)
#apply_gradient_op = tf.train.AdamOptimizer(self.init_learning_rate).minimize(self.loss)
self.train = apply_gradient_op
return apply_gradient_op