-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdcnV2.py
182 lines (152 loc) · 9.32 KB
/
dcnV2.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
import tensorflow as tf
import tensorflow.contrib as tc
def DeformableConv2D(input,
output_dims,
kernel_size,
stride,
idx,
saperable=False,
activation_fn=None,
normalizer_fn=None,
normalizer_params=None,
biases_initializer=None
):
with tf.variable_scope('DeformableConv2D'+'_'+idx, reuse=tf.AUTO_REUSE):
if saperable:
offsets = tc.layers.separable_conv2d(input, None, 3, 1,
stride=1,
activation_fn=tf.nn.relu6,
normalizer_fn=None, normalizer_params=None,
rate=1,
scope='offset_conv_depthwise')
offsets = tc.layers.conv2d(offsets, kernel_size * kernel_size * 2, 1, activation_fn=tf.nn.sigmoid,
normalizer_fn=None, normalizer_params=None,
scope='offset_conv_pointwise')
modulation = tc.layers.separable_conv2d(input, None, 3, 1,
stride=1,
activation_fn=tf.nn.relu6,
normalizer_fn=None, normalizer_params=None,
rate=1,
scope='weights_conv_depthwise')
modulation = tc.layers.conv2d(modulation, kernel_size * kernel_size, 1, activation_fn=tf.nn.sigmoid,
normalizer_fn=None, normalizer_params=None,
scope='weights_conv_pointwise')
else:
offsets = tc.layers.conv2d(input,
kernel_size*kernel_size*2,
kernel_size,
stride=stride,
activation_fn=None,
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
scope='offset_conv')
modulation = tc.layers.conv2d(input,
kernel_size*kernel_size,
kernel_size,
stride=stride,
activation_fn=tf.nn.sigmoid,
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
scope='weights_conv')
input_shape = tf.shape(input)
# [b, h, w, 2c] -> [b*c, h, w, 2]
offsets = tf.transpose(offsets, [0, 3, 1, 2])
offsets = tf.reshape(offsets, [offsets.shape[0], -1, 2, offsets.shape[2], offsets.shape[3]])
offsets = tf.transpose(offsets, [0, 1, 3, 4, 2])
offsets = tf.reshape(offsets, [-1, offsets.shape[2], offsets.shape[3], 2])
# [b, h, w, c] -> [b*c, h, w]
modulation = tf.transpose(modulation, [0, 3, 1, 2])
modulation = tf.reshape(modulation, [-1, modulation.shape[2], modulation.shape[3]])
grid_t = tf.reshape(tf.range(input_shape[1]), [-1, 1])
grid_l = tf.reshape(tf.range(input_shape[2]), [1, -1])
grid_b = tf.reshape(tf.range(input_shape[0]), [-1,1, 1, 1, 1])
grid_bk = tf.tile(grid_b, [1,kernel_size * kernel_size, 1, 1, 1])
grid_bk = tf.reshape(grid_bk, [-1,1,1, 1])
grid_t = tf.tile(grid_t, [1, input_shape[2]])
grid_l = tf.tile(grid_l, [input_shape[1], 1])
grid_tl = tf.stack([grid_t, grid_l], axis=-1)
grid_tl = tf.tile(tf.expand_dims(grid_tl, axis=0), [offsets.shape[0], 1, 1, 1])
coordi_yx = tf.cast(grid_tl, dtype=tf.float32) + offsets
coordi_y_clip = tf.clip_by_value(coordi_yx[..., 0],
clip_value_min=0.,
clip_value_max=tf.cast(input_shape[1], dtype=tf.float32) - 1.)
coordi_x_clip = tf.clip_by_value(coordi_yx[..., 1],
clip_value_min=0.,
clip_value_max=tf.cast(input_shape[2], dtype=tf.float32) - 1.)
coordi_yx = tf.stack([coordi_y_clip, coordi_x_clip], axis=-1)
grid_bk = tf.cast(tf.tile(grid_bk, [1, input_shape[1], input_shape[2], 1]), dtype=tf.float32)
coordi_bkyx = tf.concat([grid_bk, coordi_yx], axis=-1)
coordi_bktl = tf.cast(tf.floor(coordi_bkyx), dtype=tf.int32)
coordi_bktl = tf.reshape(coordi_bktl, [-1, 3])
cell = tf.reshape(tf.convert_to_tensor([0, 0, 1]), [1, -1])
cell_tile1 = tf.tile(cell, [coordi_bktl.shape[0], 1])
cell2 = tf.reshape(tf.convert_to_tensor([0, 1, 0]), [1, -1])
cell_tile2 = tf.tile(cell2, [coordi_bktl.shape[0], 1])
coordi_bktr = coordi_bktl + cell_tile1
coordi_bkbl = coordi_bktl + cell_tile2
coordi_bkbr = coordi_bktl + cell_tile1 + cell_tile2
var_bktlc = tf.gather_nd(input, coordi_bktl) ###
var_tl = tf.reshape(var_bktlc, [input_shape[0] * kernel_size * kernel_size,
input_shape[1],
input_shape[2],
input_shape[3]])
# [batch, kernel*kernel, top, right, channel]
var_bktrc = tf.gather_nd(input, coordi_bktr) ###
var_tr = tf.reshape(var_bktrc, [input_shape[0] * kernel_size * kernel_size,
input_shape[1],
input_shape[2],
input_shape[3]])
# [batch, kernel*kernel, bottom, left, channel]
var_bkblc = tf.gather_nd(input, coordi_bkbl) ###
var_bl = tf.reshape(var_bkblc, [input_shape[0] * kernel_size * kernel_size,
input_shape[1],
input_shape[2],
input_shape[3]])
# [batch, kernel*kernel, bottom, right, channel]
var_bkbrc = tf.gather_nd(input, coordi_bkbr) ###
var_br = tf.reshape(var_bkbrc, [input_shape[0] * kernel_size * kernel_size,
input_shape[1],
input_shape[2],
input_shape[3]])
offset_y = tf.expand_dims(coordi_yx[..., 0] - tf.floor(coordi_yx[..., 0]), axis=-1)
offset_x = tf.expand_dims(coordi_yx[..., 1] - tf.floor(coordi_yx[..., 1]), axis=-1)
offset_y = tf.tile(offset_y, [1, 1, 1, input_shape[3]])
offset_x = tf.tile(offset_x, [1, 1, 1, input_shape[3]])
var_top = var_tl + (var_tr - var_tl) * offset_x
var_bottom = var_bl + (var_br - var_bl) * offset_x
var_final = var_top + (var_bottom - var_top) * offset_y
modulation = tf.expand_dims(modulation, axis=-1) ###
var_final = var_final * modulation ###
var_final = tf.reshape(var_final,
[input_shape[0], kernel_size,kernel_size, input_shape[1], input_shape[2],
input_shape[3]])
var_final = tf.transpose(var_final, [0, 3, 1, 4, 2, 5])
var_final = tf.reshape(var_final,
[input_shape[0], kernel_size * input_shape[1], kernel_size * input_shape[2],
input_shape[3]])
if saperable:
var_final = tc.layers.separable_conv2d(var_final, None, kernel_size, 1,
stride=kernel_size,
activation_fn=tf.nn.relu6,
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
padding='VALID',
rate=1,
scope='output_conv_depthwise')
var_final = tc.layers.conv2d(var_final, output_dims, 1,
activation_fn=tf.nn.relu6,
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
scope='output_conv_pointwise')
else:
var_final = tc.layers.conv2d(var_final,
output_dims,
kernel_size,
stride=kernel_size,
padding='VALID',
activation_fn=activation_fn,
normalizer_fn=normalizer_fn,
biases_initializer=biases_initializer,
normalizer_params=normalizer_params,
scope='output_conv')
return var_final