-
Notifications
You must be signed in to change notification settings - Fork 57
/
flownet.py
237 lines (194 loc) · 10.7 KB
/
flownet.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
import tensorflow as tf
import tensorflow.contrib.slim as slim
import tensorflow.contrib.layers as layers
from ..ops import correlation
from .image_warp import image_warp
from .flow_util import flow_to_color
FLOW_SCALE = 5.0
def flownet(im1, im2, flownet_spec='S', full_resolution=False, train_all=False,
backward_flow=False):
num_batch, height, width, _ = tf.unstack(tf.shape(im1))
flownet_num = len(flownet_spec)
assert flownet_num > 0
flows_fw = []
flows_bw = []
for i, name in enumerate(flownet_spec):
assert name in ('C', 'c', 'S', 's')
channel_mult = 1 if name in ('C', 'S') else 3 / 8
full_res = full_resolution and i == flownet_num - 1
def scoped_block():
if name.lower() == 'c':
assert i == 0, 'FlowNetS must be used for refinement networks'
with tf.variable_scope('flownet_c_features'):
_, conv2_a, conv3_a = flownet_c_features(im1, channel_mult=channel_mult)
_, conv2_b, conv3_b = flownet_c_features(im2, channel_mult=channel_mult, reuse=True)
with tf.variable_scope('flownet_c') as scope:
flow_fw = flownet_c(conv3_a, conv3_b, conv2_a,
full_res=full_res,
channel_mult=channel_mult)
flows_fw.append(flow_fw)
if backward_flow:
scope.reuse_variables()
flow_bw = flownet_c(conv3_b, conv3_a, conv2_b,
full_res=full_res,
channel_mult=channel_mult)
flows_bw.append(flow_bw)
elif name.lower() == 's':
def _flownet_s(im1, im2, flow=None):
if flow is not None:
flow = tf.image.resize_bilinear(flow, [height, width]) * 4 * FLOW_SCALE
warp = image_warp(im2, flow)
diff = tf.abs(warp - im1)
if not train_all:
flow = tf.stop_gradient(flow)
warp = tf.stop_gradient(warp)
diff = tf.stop_gradient(diff)
inputs = tf.concat([im1, im2, flow, warp, diff], axis=3)
inputs = tf.reshape(inputs, [num_batch, height, width, 14])
else:
inputs = tf.concat([im1, im2], 3)
return flownet_s(inputs,
full_res=full_res,
channel_mult=channel_mult)
stacked = len(flows_fw) > 0
with tf.variable_scope('flownet_s') as scope:
flow_fw = _flownet_s(im1, im2, flows_fw[-1][0] if stacked else None)
flows_fw.append(flow_fw)
if backward_flow:
scope.reuse_variables()
flow_bw = _flownet_s(im2, im1, flows_bw[-1][0] if stacked else None)
flows_bw.append(flow_bw)
if i > 0:
scope_name = "stack_{}_flownet".format(i)
with tf.variable_scope(scope_name):
scoped_block()
else:
scoped_block()
if backward_flow:
return flows_fw, flows_bw
return flows_fw
def _leaky_relu(x):
with tf.variable_scope('leaky_relu'):
return tf.maximum(0.1 * x, x)
def _flownet_upconv(conv6_1, conv5_1, conv4_1, conv3_1, conv2, conv1=None, inputs=None,
channel_mult=1, full_res=False, channels=2):
m = channel_mult
flow6 = slim.conv2d(conv6_1, channels, 3, scope='flow6',
activation_fn=None)
deconv5 = slim.conv2d_transpose(conv6_1, int(512 * m), 4, stride=2,
scope='deconv5')
flow6_up5 = slim.conv2d_transpose(flow6, channels, 4, stride=2,
scope='flow6_up5',
activation_fn=None)
concat5 = tf.concat([conv5_1, deconv5, flow6_up5], 1)
flow5 = slim.conv2d(concat5, channels, 3, scope='flow5',
activation_fn=None)
deconv4 = slim.conv2d_transpose(concat5, int(256 * m), 4, stride=2,
scope='deconv4')
flow5_up4 = slim.conv2d_transpose(flow5, channels, 4, stride=2,
scope='flow5_up4',
activation_fn=None)
concat4 = tf.concat([conv4_1, deconv4, flow5_up4], 1)
flow4 = slim.conv2d(concat4, channels, 3, scope='flow4',
activation_fn=None)
deconv3 = slim.conv2d_transpose(concat4, int(128 * m), 4, stride=2,
scope='deconv3')
flow4_up3 = slim.conv2d_transpose(flow4, channels, 4, stride=2,
scope='flow4_up3',
activation_fn=None)
concat3 = tf.concat([conv3_1, deconv3, flow4_up3], 1)
flow3 = slim.conv2d(concat3, channels, 3, scope='flow3',
activation_fn=None)
deconv2 = slim.conv2d_transpose(concat3, int(64 * m), 4, stride=2,
scope='deconv2')
flow3_up2 = slim.conv2d_transpose(flow3, channels, 4, stride=2,
scope='flow3_up2',
activation_fn=None)
concat2 = tf.concat([conv2, deconv2, flow3_up2], 1)
flow2 = slim.conv2d(concat2, channels, 3, scope='flow2',
activation_fn=None)
flows = [flow2, flow3, flow4, flow5, flow6]
if full_res:
with tf.variable_scope('full_res'):
deconv1 = slim.conv2d_transpose(concat2, int(32 * m), 4, stride=2,
scope='deconv1')
flow2_up1 = slim.conv2d_transpose(flow2, channels, 4, stride=2,
scope='flow2_up1',
activation_fn=None)
concat1 = tf.concat([conv1, deconv1, flow2_up1], 1)
flow1 = slim.conv2d(concat1, channels, 3, scope='flow1',
activation_fn=None)
deconv0 = slim.conv2d_transpose(concat1, int(16 * m), 4, stride=2,
scope='deconv0')
flow1_up0 = slim.conv2d_transpose(flow1, channels, 4, stride=2,
scope='flow1_up0',
activation_fn=None)
concat0 = tf.concat([inputs, deconv0, flow1_up0], 1)
flow0 = slim.conv2d(concat0, channels, 3, scope='flow0',
activation_fn=None)
flows = [flow0, flow1] + flows
return flows
def nhwc_to_nchw(tensors):
return [tf.transpose(t, [0, 3, 1, 2]) for t in tensors]
def nchw_to_nhwc(tensors):
return [tf.transpose(t, [0, 2, 3, 1]) for t in tensors]
def flownet_s(inputs, channel_mult=1, full_res=False):
"""Given stacked inputs, returns flow predictions in decreasing resolution.
Uses FlowNetSimple.
"""
m = channel_mult
inputs = nhwc_to_nchw([inputs])[0]
with slim.arg_scope([slim.conv2d, slim.conv2d_transpose],
data_format='NCHW',
weights_regularizer=slim.l2_regularizer(0.0004),
weights_initializer=layers.variance_scaling_initializer(),
activation_fn=_leaky_relu):
conv1 = slim.conv2d(inputs, int(64 * m), 7, stride=2, scope='conv1')
conv2 = slim.conv2d(conv1, int(128 * m), 5, stride=2, scope='conv2')
conv3 = slim.conv2d(conv2, int(256 * m), 5, stride=2, scope='conv3')
conv3_1 = slim.conv2d(conv3, int(256 * m), 3, stride=1, scope='conv3_1')
conv4 = slim.conv2d(conv3_1, int(512 * m), 3, stride=2, scope='conv4')
conv4_1 = slim.conv2d(conv4, int(512 * m), 3, stride=1, scope='conv4_1')
conv5 = slim.conv2d(conv4_1, int(512 * m), 3, stride=2, scope='conv5')
conv5_1 = slim.conv2d(conv5, int(512 * m), 3, stride=1, scope='conv5_1')
conv6 = slim.conv2d(conv5_1, int(1024 * m), 3, stride=2, scope='conv6')
conv6_1 = slim.conv2d(conv6, int(1024 * m), 3, stride=1, scope='conv6_1')
res = _flownet_upconv(conv6_1, conv5_1, conv4_1, conv3_1, conv2, conv1, inputs,
channel_mult=channel_mult, full_res=full_res)
return nchw_to_nhwc(res)
def flownet_c_features(im, channel_mult=1, reuse=None):
m = channel_mult
im = nhwc_to_nchw([im])[0]
with slim.arg_scope([slim.conv2d],
data_format='NCHW',
weights_regularizer=slim.l2_regularizer(0.0004),
weights_initializer=layers.variance_scaling_initializer(),
activation_fn=_leaky_relu):
conv1 = slim.conv2d(im, int(64 * m), 7, stride=2, scope='conv1', reuse=reuse)
conv2 = slim.conv2d(conv1, int(128 * m), 5, stride=2, scope='conv2', reuse=reuse)
conv3 = slim.conv2d(conv2, int(256 * m), 5, stride=2, scope='conv3', reuse=reuse)
return conv1, conv2, conv3
def flownet_c(conv3_a, conv3_b, conv2_a, channel_mult=1, full_res=False):
"""Given two images, returns flow predictions in decreasing resolution.
Uses FlowNetCorr.
"""
m = channel_mult
with slim.arg_scope([slim.conv2d, slim.conv2d_transpose],
data_format='NCHW',
weights_regularizer=slim.l2_regularizer(0.0004),
weights_initializer=layers.variance_scaling_initializer(),
activation_fn=_leaky_relu):
corr = correlation(conv3_a, conv3_b,
pad=20, kernel_size=1, max_displacement=20, stride_1=1, stride_2=2)
conv_redir = slim.conv2d(conv3_a, int(32 * m), 1, stride=1, scope='conv_redir')
conv3_1 = slim.conv2d(tf.concat([conv_redir, corr], 1), int(256 * m), 3,
stride=1, scope='conv3_1')
conv4 = slim.conv2d(conv3_1, int(512 * m), 3, stride=2, scope='conv4')
conv4_1 = slim.conv2d(conv4, int(512 * m), 3, stride=1, scope='conv4_1')
conv5 = slim.conv2d(conv4_1, int(512 * m), 3, stride=2, scope='conv5')
conv5_1 = slim.conv2d(conv5, int(512 * m), 3, stride=1, scope='conv5_1')
conv6 = slim.conv2d(conv5_1, int(1024 * m), 3, stride=2, scope='conv6')
conv6_1 = slim.conv2d(conv6, int(1024 * m), 3, stride=1, scope='conv6_1')
res = _flownet_upconv(conv6_1, conv5_1, conv4_1, conv3_1, conv2_a,
channel_mult=channel_mult, full_res=full_res)
return nchw_to_nhwc(res)