forked from rogerhcheng/LiteFlowNet2-TF2
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
312 lines (252 loc) · 16.8 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
from warp import *
import tensorflow as tf
import tensorflow_addons as tfa
class LiteFlowNet2():
def __init__(self, isSintel = True):
self.dblBackward = [0.0, 0.0, 0.0, 5.0, 2.5, 1.25, 0.625]
self.isSintel = isSintel
def feature_extractor(self):
lrelu = lambda x: tf.nn.leaky_relu(x, 0.1)
# module one
m1 = tf.keras.Sequential()
m1.add(tf.keras.layers.Conv2D(filters=32, kernel_size=7, activation=lrelu, padding='SAME'))
# module two
m2 = tf.keras.Sequential()
m2.add(tf.keras.layers.ZeroPadding2D(padding=(1, 1)))
m2.add(tf.keras.layers.Conv2D(filters=32, kernel_size=3, strides=2, activation=lrelu, padding='valid'))
m2.add(tf.keras.layers.Conv2D(filters=32, kernel_size=3, activation=lrelu, padding='SAME'))
m2.add(tf.keras.layers.Conv2D(filters=32, kernel_size=3, activation=lrelu, padding='SAME'))
# module three
m3 = tf.keras.Sequential()
m3.add(tf.keras.layers.ZeroPadding2D(padding=(1, 1)))
m3.add(tf.keras.layers.Conv2D(filters=64, kernel_size=3, strides=2, activation=lrelu, padding='valid'))
m3.add(tf.keras.layers.Conv2D(filters=64, kernel_size=3, activation=lrelu, padding='SAME'))
# module four
m4 = tf.keras.Sequential()
m4.add(tf.keras.layers.ZeroPadding2D(padding=(1, 1)))
m4.add(tf.keras.layers.Conv2D(filters=96, kernel_size=3, strides=2, activation=lrelu, padding='valid'))
m4.add(tf.keras.layers.Conv2D(filters=96, kernel_size=3, activation=lrelu, padding='SAME'))
# module five
m5 = tf.keras.Sequential()
m5.add(tf.keras.layers.ZeroPadding2D(padding=(1, 1)))
m5.add(tf.keras.layers.Conv2D(filters=128, kernel_size=3, strides=2, activation=lrelu, padding='valid'))
# module six
m6 = tf.keras.Sequential()
m6.add(tf.keras.layers.ZeroPadding2D(padding=(1, 1)))
m6.add(tf.keras.layers.Conv2D(filters=192, kernel_size=3, strides=2, activation=lrelu, padding='valid'))
return [m1, m2, m3, m4, m5, m6]
def group_upconv(self, in1, groups, name, with_bias=False):
# keras don't have an easy way of group conv so use old way
with tf.compat.v1.variable_scope('flownet'):
with tf.compat.v1.variable_scope(name):
filterc = tf.compat.v1.get_variable('filter_w', shape=[4, 4, 1, groups], dtype=tf.float32)
shp = tf.shape(in1)
output_shape = (shp[0], shp[1] * 2, shp[2] * 2, shp[3])
out = tf.nn.conv2d_transpose(in1, filterc, output_shape, strides=[1, 2, 2, 1])
if with_bias:
bias = tf.compat.v1.get_variable('filter_b', shape=[groups], dtype=tf.float32)
out = tf.nn.bias_add(out, bias)
return out
def matching(self, tensor_features1, tensor_features2, tensorFlow, int_level, name):
with tf.name_scope(name):
lrelu = lambda x: tf.nn.leaky_relu(x, 0.1)
def module_feat():
return tf.keras.Sequential()
def module_upcorr(x):
return self.group_upconv(x, 49, name + '/moduleUpcorr')
def module_upflow(x):
return self.group_upconv(x, 2, name + '/moduleUpflow')
def module_main(x):
kernel_size = [0, 0, 0, 5, 5, 3, 3][int_level]
if kernel_size == 0:
raise ValueError('Should not be in level %i in Matching layer!' % int_level)
with tf.name_scope('module_main'):
conv1 = tf.keras.layers.Conv2D(filters=128, kernel_size=3, activation=lrelu, padding='SAME')(x)
conv2 = tf.keras.layers.Conv2D(filters=128, kernel_size=3, activation=lrelu, padding='SAME')(conv1)
conv3 = tf.keras.layers.Conv2D(filters=96, kernel_size=3, activation=lrelu, padding='SAME')(conv2)
conv4 = tf.keras.layers.Conv2D(filters=64, kernel_size=3, activation=lrelu, padding='SAME')(conv3)
conv5 = tf.keras.layers.Conv2D(filters=32, kernel_size=3, activation=lrelu, padding='SAME')(conv4)
conv6 = tf.keras.layers.Conv2D(filters=2, kernel_size=kernel_size, activation=None, padding='SAME')(conv5)
return conv6
if self.dblBackward[int_level] == 0.0:
raise ValueError('Should not be in level %i in Matching layer!' % int_level)
with tf.name_scope('module_feat'):
m_feat = module_feat()
tensor_features1 = m_feat(tensor_features1)
tensor_features2 = m_feat(tensor_features2)
if tensorFlow is not None:
tensorFlow = module_upflow(tensorFlow)
# warp features
tensor_features2 = tf_warp(tensor_features2, tensorFlow * self.dblBackward[int_level])
if int_level >= 4:
corr = tfa.layers.optical_flow.CorrelationCost(1, 3, 1, 1, 3, 'channels_last')([tensor_features1, tensor_features2])
corr = lrelu(corr)
else:
corr = tfa.layers.optical_flow.CorrelationCost(1, 6, 2, 2, 6, 'channels_last')([tensor_features1, tensor_features2])
corr = lrelu(module_upcorr(corr))
# hack cuz corr cost lost last dimension
corr.set_shape([None, None, None, 49])
return (tensorFlow if tensorFlow is not None else 0.0) + module_main(corr)
def subpixel(self, tensor_features1, tensor_features2, tensorFlow, int_level, name='subpixel'):
with tf.name_scope(name):
lrelu = lambda x: tf.nn.leaky_relu(x, 0.1)
def module_feat():
return tf.keras.Sequential()
def module_main(x):
kernel_size = [0, 0, 0, 5, 5, 3, 3][int_level]
if kernel_size == 0:
raise ValueError('Should not be in level %i in Subpixel layer!' % int_level)
with tf.name_scope("module_main"):
conv1 = tf.keras.layers.Conv2D(filters=128, kernel_size=3, activation=lrelu, padding='SAME')(x)
conv2 = tf.keras.layers.Conv2D(filters=128, kernel_size=3, activation=lrelu, padding='SAME')(conv1)
conv3 = tf.keras.layers.Conv2D(filters=96, kernel_size=3, activation=lrelu, padding='SAME')(conv2)
conv4 = tf.keras.layers.Conv2D(filters=64, kernel_size=3, activation=lrelu, padding='SAME')(conv3)
conv5 = tf.keras.layers.Conv2D(filters=32, kernel_size=3, activation=lrelu, padding='SAME')(conv4)
conv6 = tf.keras.layers.Conv2D(filters=2, kernel_size=kernel_size, activation=None, padding='SAME')(conv5)
if int_level > 3 or self.isSintel:
return conv6
else:
return [conv6, conv5]
with tf.name_scope('module_feat'):
mfeat = module_feat()
tensor_features1 = mfeat(tensor_features1)
tensor_features2 = mfeat(tensor_features2)
if self.dblBackward[int_level] == 0:
raise ValueError('Should not be in level %i in Subpixel layer!' % int_level)
tensorFlow1 = tensorFlow * self.dblBackward[int_level]
tensor_features2 = tf_warp(tensor_features2, tensorFlow1)
tens_flow = tf.concat([tensor_features1, tensor_features2, tensorFlow], -1)
if int_level > 3 or self.isSintel:
return (tensorFlow if tensorFlow is not None else 0.0) + module_main(tens_flow)
else:
main_output = module_main(tens_flow)
return [(tensorFlow if tensorFlow is not None else 0.0) + main_output[0], main_output[1]]
def regularization(self, tensor1, tensor2, tensor_features1, tensorFlow, int_level, name='module_regularization', intermediate_tensor=None):
with tf.name_scope(name):
lrelu = lambda x: tf.nn.leaky_relu(x, 0.1)
int_unfold = [0, 0, 7, 5, 5, 3, 3][int_level]
if int_unfold == 0:
raise ValueError('Should not be in level %i in Regularization layer!' % int_level)
def module_feat(x):
with tf.name_scope('module_feat'):
if int_level == 3 or int_level == 4:
return tf.keras.layers.Conv2D(filters=128, kernel_size=1, activation=lrelu, padding='valid')(x)
else:
return x
moduleScale = lambda x: tf.keras.layers.Conv2D(filters=1, kernel_size=1, activation=None, padding='valid')(x)
def module_main(x):
if int_level > 2:
with tf.name_scope('module_main'):
conv1 = tf.keras.layers.Conv2D(filters=128, kernel_size=3, activation=lrelu, padding='SAME')(x)
conv2 = tf.keras.layers.Conv2D(filters=128, kernel_size=3, activation=lrelu, padding='SAME')(conv1)
conv3 = tf.keras.layers.Conv2D(filters=64, kernel_size=3, activation=lrelu, padding='SAME')(conv2)
conv4 = tf.keras.layers.Conv2D(filters=64, kernel_size=3, activation=lrelu, padding='SAME')(conv3)
conv5 = tf.keras.layers.Conv2D(filters=32, kernel_size=3, activation=lrelu, padding='SAME')(conv4)
conv6 = tf.keras.layers.Conv2D(filters=32, kernel_size=3, activation=lrelu, padding='SAME')(conv5)
return conv6
else:
with tf.name_scope('moduleUpflowR'):
conv6 = self.group_upconv(x, 32, name + '/moduleUpflowR', True)
return conv6
def module_dist(x):
with tf.name_scope('module_dist'):
kernel_size = [0, 0, 7, 5, 5, 3, 3][int_level]
out_channels = [0, 0, 49, 25, 25, 9, 9][int_level]
if kernel_size == 0 or out_channels == 0:
raise ValueError('Should not be in level %i in Regularization layer!' % int_level)
if int_level >= 5:
return tf.keras.layers.Conv2D(filters=out_channels, kernel_size=kernel_size, padding='SAME', activation=None,)(x)
else:
x = tf.keras.layers.Conv2D(filters=out_channels, kernel_size=(kernel_size, 1), activation=None,
padding='same')(x)
x = tf.keras.layers.Conv2D(filters=out_channels, kernel_size=(1, kernel_size), activation=None,
padding='same')(x)
return x
if int_level > 2:
if self.dblBackward[int_level] == 0:
raise ValueError('Should not be in level %i in Regularization layer!' % int_level)
tensor_diff = tf.sqrt(tf.reduce_sum(tf.square(tensor1 - tf_warp(tensor2, tensorFlow * self.dblBackward[int_level])),
axis=3, keepdims=True))
feat = module_feat(tensor_features1)
intermediate_tensor = module_main(tf.concat([tensor_diff,
tensorFlow - tf.reduce_mean(tensorFlow, keepdims=True,
axis=[1, 2]),
feat], 3))
else:
intermediate_tensor = self.group_upconv(intermediate_tensor, 32, name + '/moduleUpflowR', True)
tensor_dist = module_dist(intermediate_tensor)
tensor_dist = -tf.square(tensor_dist)
tensor_dist = tf.exp(tensor_dist - tf.reduce_max(tensor_dist, axis=3, keepdims=True))
tensor_div = 1. / tf.reduce_sum(tensor_dist, -1, keepdims=True)
tensorScale = []
moduleScaleNames = ['moduleScaleX', 'moduleScaleY']
indices = [0, 1, 2]
for i in range(len(moduleScaleNames)):
with tf.name_scope(moduleScaleNames[i]):
tensorScale.append(moduleScale(tensor_dist *
tf.image.extract_patches(tensorFlow[..., indices[i]:indices[i+1]],
[1, int_unfold, int_unfold, 1],
[1, 1, 1, 1],
[1, 1, 1, 1],
"SAME")))
if int_level != 3 or self.isSintel:
return tf.concat([tensorScale[0] * tensor_div, tensorScale[1] * tensor_div], -1)
else:
return [tf.concat([tensorScale[0] * tensor_div, tensorScale[1] * tensor_div], -1), intermediate_tensor]
def correct_pan(self, x):
with tf.name_scope('correct_pan'):
lrelu = lambda x: tf.nn.leaky_relu(x, 0.1)
conv1 = tf.keras.layers.Conv2D(filters=64, kernel_size=3, activation=lrelu, padding='SAME')(x)
conv2 = tf.keras.layers.Conv2D(filters=32, kernel_size=3, activation=lrelu, padding='SAME')(conv1)
conv3 = tf.keras.layers.Conv2D(filters=1, kernel_size=1, activation=None, padding='valid')(conv2)
return tf.nn.tanh(conv3)
def module_chromas(self, x):
with tf.name_scope('module_chromas'):
lrelu = lambda x: tf.nn.leaky_relu(x, 0.1)
conv1 = tf.keras.layers.Conv2D(filters=32, kernel_size=3, activation=lrelu, padding='SAME')(x)
conv2 = tf.keras.layers.Conv2D(filters=2, kernel_size=1, activation=None, padding='valid')(conv1)
return conv2
def __call__(self, tensor1, tensor2, scope='flownet'):
tf.keras.backend.set_floatx('float32')
with tf.name_scope(scope):
tensor1_norm = tensor1 - [[[[0.411618, 0.434631, 0.454253]]]]
tensor2_norm = tensor2 - [[[[0.410782, 0.433645, 0.452793]]]]
m1, m2, m3, m4, m5, m6 = self.feature_extractor()
def shared_feat_modules(x):
with tf.name_scope('feature_extractor'):
t1 = m1(x)
t2 = m2(t1)
t3 = m3(t2)
t4 = m4(t3)
t5 = m5(t4)
t6 = m6(t5)
return [t1, t2, t3, t4, t5, t6]
# Extract features
tensor_feat1 = shared_feat_modules(tensor1_norm)
tensor_feat2 = shared_feat_modules(tensor2_norm)
tensor1 = [tensor1_norm]
tensor2 = [tensor2_norm]
for i in [2, 3, 4, 5]:
tensor1.append(tf.image.resize(tensor1[-1], tf.shape(tensor_feat1[i])[1:3]))
tensor2.append(tf.image.resize(tensor2[-1], tf.shape(tensor_feat2[i])[1:3]))
# Main loop
flow = None
lvls = [2, 3, 4, 5, 6]
for i in [-1, -2, -3, -4]:
flow = self.matching(tensor_feat1[i], tensor_feat2[i], flow, lvls[i], name='matching_%i' % abs(i))
if lvls[i] == 3 and not self.isSintel:
[flow, intermediate_S_tensor] = self.subpixel(tensor_feat1[i], tensor_feat2[i], flow, lvls[i], name='subpixel_%i' % abs(i))
[flow, intermediate_R_tensor] = self.regularization(tensor1[i], tensor2[i], tensor_feat1[i], flow, lvls[i], name='regularization_%i' % abs(i))
else:
flow = self.subpixel(tensor_feat1[i], tensor_feat2[i], flow, lvls[i], name='subpixel_%i' % abs(i))
flow = self.regularization(tensor1[i], tensor2[i], tensor_feat1[i], flow, lvls[i], name='regularization_%i' % abs(i))
# Go through extra layers in Kitti model
if not self.isSintel:
intermediate_S_tensor = self.group_upconv(intermediate_S_tensor, 32, 'subpixel_5/moduleUpflowS', True)
with tf.name_scope("subpixel_5/module_main"):
intermediate_S_tensor = tf.keras.layers.Conv2D(filters=2, kernel_size=7, activation=None, padding='SAME')(intermediate_S_tensor)
flow = self.group_upconv(flow, 2, 'matching_5/moduleUpflow', False) + intermediate_S_tensor
i = -5
flow = self.regularization(None, None, tensor_feat1[i], flow, lvls[i], name='regularization_%i' % abs(i), intermediate_tensor=intermediate_R_tensor)
flowr = tf.image.resize(flow, tf.shape(tensor1_norm)[1:3])
return flowr * 20.0