-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathattention_module.py
235 lines (213 loc) · 7.33 KB
/
attention_module.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
import tensorflow as tf
from network_configure import conf_attn_same, conf_attn_up, conf_attn_down
from basic_ops import *
from attention_layer_q import self_attention as self_attention_q
from attention_layer import self_attention
"""This script defines the attention modules: same-, up-, down- global transformer operators.
Note that pre-activation is used for residual-like blocks.
"""
def same_gto(inputs, output_filters, training, dimension, name):
"""Same GTO block.
Args:
inputs: a Tensor with shape [batch, (d,) h, w, channels]
output_filters: an integer
training: a boolean for batch normalization and dropout
dimension: a string, dimension of inputs/outputs -- 2D, 3D
name: a string
Returns:
A Tensor of shape [batch, (_d,) _h, _w, output_filters]
"""
with tf.variable_scope(name):
shortcut = inputs
inputs = batch_norm(inputs, training, 'batch_norm')
inputs = relu(inputs, 'relu')
if 0:#training:
inputs_1, inputs_2 = tf.split(inputs, 2, 0)
inputs_1, q, k = self_attention(
inputs_1,
output_filters // conf_attn_same['key_ratio'],
output_filters // conf_attn_same['value_ratio'],
output_filters,
conf_attn_same['num_heads'],
training,
dimension,
'SAME',
'attention',
conf_attn_same['dropout_rate'],
conf_attn_same['use_softmax'],
conf_attn_same['use_bias'])
inputs_2, q, k = self_attention(
inputs_2,
output_filters // conf_attn_same['key_ratio'],
output_filters // conf_attn_same['value_ratio'],
output_filters,
conf_attn_same['num_heads'],
training,
dimension,
'SAME',
'attention',
conf_attn_same['dropout_rate'],
conf_attn_same['use_softmax'],
conf_attn_same['use_bias'], batch_att=False)
inputs = tf.concat([inputs_1, inputs_2], axis=0)
else:
inputs, q, k = self_attention(
inputs,
output_filters // conf_attn_same['key_ratio'],
output_filters // conf_attn_same['value_ratio'],
output_filters,
conf_attn_same['num_heads'],
training,
dimension,
'SAME',
'attention',
conf_attn_same['dropout_rate'],
conf_attn_same['use_softmax'],
conf_attn_same['use_bias'], batch_att=False)
return tf.add(shortcut, inputs), inputs, q, k
def same_gto_q(inputs, output_filters, training, dimension, name):
"""Same GTO block.
Args:
inputs: a Tensor with shape [batch, (d,) h, w, channels]
output_filters: an integer
training: a boolean for batch normalization and dropout
dimension: a string, dimension of inputs/outputs -- 2D, 3D
name: a string
Returns:
A Tensor of shape [batch, (_d,) _h, _w, output_filters]
"""
with tf.variable_scope(name):
shortcut = inputs
inputs = batch_norm(inputs, training, 'batch_norm')
inputs = relu(inputs, 'relu')
inputs, q, k = self_attention_q(
inputs,
output_filters // conf_attn_same['key_ratio'],
output_filters // conf_attn_same['value_ratio'],
output_filters,
conf_attn_same['num_heads'],
training,
dimension,
'SAME',
'attention',
conf_attn_same['dropout_rate'],
conf_attn_same['use_softmax'],
conf_attn_same['use_bias'], batch_att=False)
return tf.add(shortcut, inputs), inputs, q, k
def up_gto_v1(inputs, output_filters, training, dimension, name):
"""Up GTO block version 1."""
if dimension == '2D':
projection_shortcut = transposed_convolution_2D
elif dimension == '3D':
projection_shortcut = transposed_convolution_3D
else:
raise ValueError("Dimension (%s) must be 2D or 3D." % (dimension))
with tf.variable_scope(name):
# The projection_shortcut should come after the batch norm and ReLU.
inputs = batch_norm(inputs, training, 'batch_norm')
inputs = relu(inputs, 'relu')
shortcut = projection_shortcut(inputs, output_filters, 3, 2, False, 'projection_shortcut')
inputs, _ = self_attention(
inputs,
output_filters // conf_attn_up['key_ratio'],
output_filters // conf_attn_up['value_ratio'],
output_filters,
conf_attn_up['num_heads'],
training,
dimension,
'UP',
'attention',
conf_attn_up['dropout_rate'],
conf_attn_up['use_softmax'],
conf_attn_up['use_bias'])
return tf.add(shortcut, inputs)
def down_gto_v1(inputs, output_filters, training, dimension, name):
"""Down GTO block version 1."""
if dimension == '2D':
projection_shortcut = convolution_2D
elif dimension == '3D':
projection_shortcut = convolution_3D
else:
raise ValueError("Dimension (%s) must be 2D or 3D." % (dimension))
with tf.variable_scope(name):
# The projection_shortcut should come after the batch norm and ReLU.
inputs = batch_norm(inputs, training, 'batch_norm')
inputs = relu(inputs, 'relu')
shortcut = projection_shortcut(inputs, output_filters, 3, 2, False, 'projection_shortcut')
inputs, _ = self_attention(
inputs,
output_filters // conf_attn_down['key_ratio'],
output_filters // conf_attn_down['value_ratio'],
output_filters,
conf_attn_down['num_heads'],
training,
dimension,
'DOWN',
'attention',
conf_attn_down['dropout_rate'],
conf_attn_down['use_softmax'],
conf_attn_down['use_bias'])
return tf.add(shortcut, inputs)
def up_gto_v2(inputs, output_filters, training, dimension, name):
"""Up GTO block version 2. (Yaochen)"""
if conf_attn_up['key_ratio'] != 1:
raise ValueError("Must set key_ratio == 1!")
with tf.variable_scope(name):
inputs = batch_norm(inputs, training, 'batch_norm')
inputs = relu(inputs, 'relu')
inputs, query = self_attention(
inputs,
output_filters // conf_attn_up['key_ratio'],
output_filters // conf_attn_up['value_ratio'],
output_filters,
conf_attn_up['num_heads'],
training,
dimension,
'UP',
'attention',
conf_attn_up['dropout_rate'],
conf_attn_up['use_softmax'],
conf_attn_up['use_bias'])
return tf.add(query, inputs)
def up4_gto_v2(inputs, output_filters, training, dimension, name):
"""Up GTO block version 2. (Yaochen)"""
if conf_attn_up['key_ratio'] != 1:
raise ValueError("Must set key_ratio == 1!")
with tf.variable_scope(name):
inputs = batch_norm(inputs, training, 'batch_norm')
inputs = relu(inputs, 'relu')
inputs, query = self_attention(
inputs,
output_filters // conf_attn_up['key_ratio'],
output_filters // conf_attn_up['value_ratio'],
output_filters,
conf_attn_up['num_heads'],
training,
dimension,
'UP4',
'attention',
conf_attn_up['dropout_rate'],
conf_attn_up['use_softmax'],
conf_attn_up['use_bias'])
return tf.add(query, inputs)
def down_gto_v2(inputs, output_filters, training, dimension, name):
"""Down GTO block version 2. (Yaochen)"""
if conf_attn_up['key_ratio'] != 1:
raise ValueError("Must set key_ratio == 1!")
with tf.variable_scope(name):
inputs = batch_norm(inputs, training, 'batch_norm')
inputs = relu(inputs, 'relu')
inputs, query = self_attention(
inputs,
output_filters // conf_attn_down['key_ratio'],
output_filters // conf_attn_down['value_ratio'],
output_filters,
conf_attn_down['num_heads'],
training,
dimension,
'DOWN',
'attention',
conf_attn_down['dropout_rate'],
conf_attn_down['use_softmax'],
conf_attn_down['use_bias'])
return tf.add(query, inputs)