-
Notifications
You must be signed in to change notification settings - Fork 9
/
sig_nets.py
executable file
·318 lines (269 loc) · 13.1 KB
/
sig_nets.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
317
318
# The network design is based on Tinghui Zhou & Clement Godard's works:
# https://github.com/tinghuiz/SfMLearner/blob/master/nets.py
# https://github.com/mrharicot/monodepth/blob/master/monodepth_model.py
from __future__ import division
import tensorflow as tf
import tensorflow.contrib.slim as slim
import numpy as np
# Range of disparity/inverse depth values
DISP_SCALING_RESNET50 = 5
DISP_SCALING_VGG = 10
FLOW_SCALING = 0.1
#TODO acquire global opt configuration
gopt=None
def disp_net(opt, dispnet_inputs, is_sem_separately=False):
#TODO manually take control of batch_norm is_training flag
global gopt
gopt=opt
is_training = (opt.mode == 'train_rigid') and opt.batch_norm_is_training
if opt.new_sem_dispnet and is_sem_separately:
var_scope="depth_sem_net"
else:
var_scope="depth_net"
if opt.dispnet_encoder == 'vgg':
return build_vgg(dispnet_inputs, get_disp_vgg, is_training, var_scope)
else:
return build_resnet50(dispnet_inputs, get_disp_resnet50, is_training, var_scope)
#TODO transfer network
def seg_net(opt, segnet_inputs):
get_pred = get_seg_vgg
H = segnet_inputs.get_shape()[1].value
W = segnet_inputs.get_shape()[2].value
is_training = (opt.mode == 'train_rigid') and opt.batch_norm_is_training
batch_norm_params = {'is_training': is_training}
num_channel=0
if opt.transfer_learn_sem:
num_channel+=opt.sem_num_class
if opt.transfer_learn_ins0:
num_channel+=opt.ins_num_class
if opt.transfer_learn_ins1_edge:
num_channel+=1
with tf.variable_scope('seg_net') as sc:
with slim.arg_scope([slim.conv2d],
normalizer_fn=slim.batch_norm if opt.enable_batch_norm else None,
normalizer_params=batch_norm_params if opt.enable_batch_norm else None,
weights_regularizer=slim.l2_regularizer(0.0001),
activation_fn=tf.nn.relu):
# ENCODING
conv1 = slim.conv2d(segnet_inputs, 32, 7, 2)
conv1b = slim.conv2d(conv1, 32, 7, 1)
conv2 = slim.conv2d(conv1b, 64, 5, 2)
conv2b = slim.conv2d(conv2, 64, 5, 1)
conv3 = slim.conv2d(conv2b, 128, 3, 2)
conv3b = slim.conv2d(conv3, 128, 3, 1)
pred4 = get_pred(conv3b, num_channel)
upconv3 = upconv(conv3b, 64, 3, 2)
i3_in = tf.concat([upconv3, conv2b], axis=3)
iconv3 = slim.conv2d(i3_in, 64, 3, 1)
pred3 = get_pred(iconv3, num_channel)
pred3_up = tf.image.resize_bilinear(pred3, [np.int(H/2), np.int(W/2)])
upconv2 = upconv(iconv3, 32, 3, 2)
i2_in = tf.concat([upconv2, conv1b, pred3_up], axis=3)
iconv2 = slim.conv2d(i2_in, 32, 3, 1)
pred2 = get_pred(iconv2, num_channel)
pred2_up = tf.image.resize_bilinear(pred2, [H, W])
upconv1 = upconv(iconv2, 16, 3, 2)
i1_in = tf.concat([upconv1, pred2_up], axis=3)
iconv1 = slim.conv2d(i1_in, 16, 3, 1)
pred1 = get_pred(iconv1, num_channel)
#print(pred1.shape, pred2.shape, pred3.shape, pred4.shape)
return [pred1, pred2, pred3, pred4]
def flow_net(opt, flownet_inputs):
#TODO manually take control of batch_norm is_training flag
global gopt
gopt=opt
is_training = (opt.mode == 'train_flow') and opt.batch_norm_is_training
return build_resnet50(flownet_inputs, get_flow, is_training, 'flow_net')
def pose_net(opt, posenet_inputs, is_sem_separately=False):
#TODO manually take control of batch_norm is_training flag
global gopt
gopt=opt
if opt.new_sem_posenet and is_sem_separately:
var_scope="pose_sem_net"
else:
var_scope="pose_net"
is_training = (opt.mode == 'train_rigid') and opt.batch_norm_is_training
batch_norm_params = {'is_training': is_training}
with tf.variable_scope(var_scope) as sc:
with slim.arg_scope([slim.conv2d],
normalizer_fn=slim.batch_norm if opt.enable_batch_norm else None,
normalizer_params=batch_norm_params if opt.enable_batch_norm else None,
weights_regularizer=slim.l2_regularizer(0.0001),
activation_fn=tf.nn.relu):
conv1 = slim.conv2d(posenet_inputs, 16, 7, 2)
conv2 = slim.conv2d(conv1, 32, 5, 2)
conv3 = slim.conv2d(conv2, 64, 3, 2)
conv4 = slim.conv2d(conv3, 128, 3, 2)
conv5 = slim.conv2d(conv4, 256, 3, 2)
conv6 = slim.conv2d(conv5, 256, 3, 2)
conv7 = slim.conv2d(conv6, 256, 3, 2)
pose_pred = slim.conv2d(conv7, 6*opt.num_source, 1, 1,
normalizer_fn=None, activation_fn=None)
pose_avg = tf.reduce_mean(pose_pred, [1, 2])
pose_final = 0.01 * tf.reshape(pose_avg, [-1, opt.num_source, 6])
return pose_final
def build_vgg(inputs, get_pred, is_training, var_scope):
batch_norm_params = {'is_training': is_training}
H = inputs.get_shape()[1].value
W = inputs.get_shape()[2].value
with tf.variable_scope(var_scope) as sc:
with slim.arg_scope([slim.conv2d, slim.conv2d_transpose],
normalizer_fn=slim.batch_norm if gopt.enable_batch_norm else None,
normalizer_params=batch_norm_params if gopt.enable_batch_norm else None,
weights_regularizer=slim.l2_regularizer(0.0001),
activation_fn=tf.nn.relu):
# ENCODING
conv1 = slim.conv2d(inputs, 32, 7, 2)
conv1b = slim.conv2d(conv1, 32, 7, 1)
conv2 = slim.conv2d(conv1b, 64, 5, 2)
conv2b = slim.conv2d(conv2, 64, 5, 1)
conv3 = slim.conv2d(conv2b, 128, 3, 2)
conv3b = slim.conv2d(conv3, 128, 3, 1)
conv4 = slim.conv2d(conv3b, 256, 3, 2)
conv4b = slim.conv2d(conv4, 256, 3, 1)
conv5 = slim.conv2d(conv4b, 512, 3, 2)
conv5b = slim.conv2d(conv5, 512, 3, 1)
conv6 = slim.conv2d(conv5b, 512, 3, 2)
conv6b = slim.conv2d(conv6, 512, 3, 1)
conv7 = slim.conv2d(conv6b, 512, 3, 2)
conv7b = slim.conv2d(conv7, 512, 3, 1)
# DECODING
upconv7 = upconv(conv7b, 512, 3, 2)
# There might be dimension mismatch due to uneven down/up-sampling
upconv7 = resize_like(upconv7, conv6b)
i7_in = tf.concat([upconv7, conv6b], axis=3)
iconv7 = slim.conv2d(i7_in, 512, 3, 1)
upconv6 = upconv(iconv7, 512, 3, 2)
upconv6 = resize_like(upconv6, conv5b)
i6_in = tf.concat([upconv6, conv5b], axis=3)
iconv6 = slim.conv2d(i6_in, 512, 3, 1)
upconv5 = upconv(iconv6, 256, 3, 2)
upconv5 = resize_like(upconv5, conv4b)
i5_in = tf.concat([upconv5, conv4b], axis=3)
iconv5 = slim.conv2d(i5_in, 256, 3, 1)
upconv4 = upconv(iconv5, 128, 3, 2)
i4_in = tf.concat([upconv4, conv3b], axis=3)
iconv4 = slim.conv2d(i4_in, 128, 3, 1)
pred4 = get_pred(iconv4)
pred4_up = tf.image.resize_bilinear(pred4, [np.int(H/4), np.int(W/4)])
upconv3 = upconv(iconv4, 64, 3, 2)
i3_in = tf.concat([upconv3, conv2b, pred4_up], axis=3)
iconv3 = slim.conv2d(i3_in, 64, 3, 1)
pred3 = get_pred(iconv3)
pred3_up = tf.image.resize_bilinear(pred3, [np.int(H/2), np.int(W/2)])
upconv2 = upconv(iconv3, 32, 3, 2)
i2_in = tf.concat([upconv2, conv1b, pred3_up], axis=3)
iconv2 = slim.conv2d(i2_in, 32, 3, 1)
pred2 = get_pred(iconv2)
pred2_up = tf.image.resize_bilinear(pred2, [H, W])
upconv1 = upconv(iconv2, 16, 3, 2)
i1_in = tf.concat([upconv1, pred2_up], axis=3)
iconv1 = slim.conv2d(i1_in, 16, 3, 1)
pred1 = get_pred(iconv1)
return [pred1, pred2, pred3, pred4]
def build_resnet50(inputs, get_pred, is_training, var_scope):
batch_norm_params = {'is_training': is_training}
with tf.variable_scope(var_scope) as sc:
with slim.arg_scope([slim.conv2d, slim.conv2d_transpose],
normalizer_fn=slim.batch_norm if gopt.enable_batch_norm else None,
normalizer_params=batch_norm_params if gopt.enable_batch_norm else None,
weights_regularizer=slim.l2_regularizer(0.0001),
activation_fn=tf.nn.relu):
conv1 = conv(inputs, 64, 7, 2) # H/2 - 64D
pool1 = maxpool(conv1, 3) # H/4 - 64D
conv2 = resblock(pool1, 64, 3) # H/8 - 256D
conv3 = resblock(conv2, 128, 4) # H/16 - 512D
conv4 = resblock(conv3, 256, 6) # H/32 - 1024D
conv5 = resblock(conv4, 512, 3) # H/64 - 2048D
skip1 = conv1
skip2 = pool1
skip3 = conv2
skip4 = conv3
skip5 = conv4
# DECODING
upconv6 = upconv(conv5, 512, 3, 2) #H/32
upconv6 = resize_like(upconv6, skip5)
concat6 = tf.concat([upconv6, skip5], 3)
iconv6 = conv(concat6, 512, 3, 1)
upconv5 = upconv(iconv6, 256, 3, 2) #H/16
upconv5 = resize_like(upconv5, skip4)
concat5 = tf.concat([upconv5, skip4], 3)
iconv5 = conv(concat5, 256, 3, 1)
upconv4 = upconv(iconv5, 128, 3, 2) #H/8
upconv4 = resize_like(upconv4, skip3)
concat4 = tf.concat([upconv4, skip3], 3)
iconv4 = conv(concat4, 128, 3, 1)
pred4 = get_pred(iconv4)
upred4 = upsample_nn(pred4, 2)
upconv3 = upconv(iconv4, 64, 3, 2) #H/4
concat3 = tf.concat([upconv3, skip2, upred4], 3)
iconv3 = conv(concat3, 64, 3, 1)
pred3 = get_pred(iconv3)
upred3 = upsample_nn(pred3, 2)
upconv2 = upconv(iconv3, 32, 3, 2) #H/2
concat2 = tf.concat([upconv2, skip1, upred3], 3)
iconv2 = conv(concat2, 32, 3, 1)
pred2 = get_pred(iconv2)
upred2 = upsample_nn(pred2, 2)
upconv1 = upconv(iconv2, 16, 3, 2) #H
concat1 = tf.concat([upconv1, upred2], 3)
iconv1 = conv(concat1, 16, 3, 1)
pred1 = get_pred(iconv1)
return [pred1, pred2, pred3, pred4]
#TODO changed
def conv(x, num_out_layers, kernel_size, stride, activation_fn=tf.nn.relu, normalizer_fn=slim.batch_norm):
p = np.floor((kernel_size - 1) / 2).astype(np.int32)
p_x = tf.pad(x, [[0, 0], [p, p], [p, p], [0, 0]])
normalizer_fn = normalizer_fn if gopt.enable_batch_norm else None
return slim.conv2d(p_x, num_out_layers, kernel_size, stride, 'VALID', activation_fn=activation_fn, normalizer_fn=normalizer_fn)
def maxpool(x, kernel_size):
p = np.floor((kernel_size - 1) / 2).astype(np.int32)
p_x = tf.pad(x, [[0, 0], [p, p], [p, p], [0, 0]])
return slim.max_pool2d(p_x, kernel_size)
def get_disp_vgg(x):
disp = DISP_SCALING_VGG * slim.conv2d(x, 1, 3, 1, activation_fn=tf.nn.sigmoid, normalizer_fn=None) + 0.01
return disp
def get_disp_resnet50(x):
disp = DISP_SCALING_RESNET50 * conv(x, 1, 3, 1, activation_fn=tf.nn.sigmoid, normalizer_fn=None) + 0.01
return disp
#TODO
def get_seg_vgg(x,channel):
seg = slim.conv2d(x, channel, 3, 1, activation_fn=tf.nn.sigmoid, normalizer_fn=None)
return seg
def get_flow(x):
# Output flow value is normalized by image height/width
flow = FLOW_SCALING * slim.conv2d(x, 2, 3, 1, activation_fn=None, normalizer_fn=None)
return flow
def resize_like(inputs, ref):
iH, iW = inputs.get_shape()[1], inputs.get_shape()[2]
rH, rW = ref.get_shape()[1], ref.get_shape()[2]
if iH == rH and iW == rW:
return inputs
return tf.image.resize_nearest_neighbor(inputs, [rH.value, rW.value])
def upsample_nn(x, ratio):
h = x.get_shape()[1].value
w = x.get_shape()[2].value
return tf.image.resize_nearest_neighbor(x, [h * ratio, w * ratio])
def upconv(x, num_out_layers, kernel_size, scale):
upsample = upsample_nn(x, scale)
cnv = conv(upsample, num_out_layers, kernel_size, 1)
return cnv
def resconv(x, num_layers, stride):
# Actually here exists a bug: tf.shape(x)[3] != num_layers is always true,
# but we preserve it here for consistency with Godard's implementation.
do_proj = tf.shape(x)[3] != num_layers or stride == 2
shortcut = []
conv1 = conv(x, num_layers, 1, 1)
conv2 = conv(conv1, num_layers, 3, stride)
conv3 = conv(conv2, 4 * num_layers, 1, 1, None)
if do_proj:
shortcut = conv(x, 4 * num_layers, 1, stride, None)
else:
shortcut = x
return tf.nn.relu(conv3 + shortcut)
def resblock(x, num_layers, num_blocks):
out = x
for i in range(num_blocks - 1):
out = resconv(out, num_layers, 1)
out = resconv(out, num_layers, 2)
return out