-
Notifications
You must be signed in to change notification settings - Fork 0
/
layer.py
executable file
·205 lines (179 loc) · 8.93 KB
/
layer.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
import tensorflow as tf
import keras
from keras import backend as K
from utils import _conv2d_wrapper, _get_weights_wrapper
def softmax(x, axis=-1):
ex = K.exp(x - K.max(x, axis=axis, keepdims=True))
return ex/K.sum(ex, axis=axis, keepdims=True)
def squash_v1(x, axis=-1):
s_squared_norm = K.sum(K.square(x), axis, keepdims=True) + K.epsilon()
scale = K.sqrt(s_squared_norm)/ (0.5 + s_squared_norm)
return scale * x
def squash_v0(s, axis=-1, epsilon=1e-7, name=None):
s_squared_norm = K.sum(K.square(s), axis, keepdims=True) + K.epsilon()
safe_norm = K.sqrt(s_squared_norm)
scale = 1 - tf.exp(-safe_norm)
return scale * s / safe_norm
def routing(u_hat_vecs, beta_a, iterations, output_capsule_num, i_activations):
b = keras.backend.zeros_like(u_hat_vecs[:,:,:,0])
if i_activations is not None:
i_activations = i_activations[...,tf.newaxis]
for i in range(iterations):
if False:
leak = tf.zeros_like(b, optimize=True)
leak = tf.reduce_sum(leak, axis=1, keep_dims=True)
leaky_logits = tf.concat([leak, b], axis=1)
leaky_routing = tf.nn.softmax(leaky_logits, dim=1)
c = tf.split(leaky_routing, [1, output_capsule_num], axis=1)[1]
else:
c = softmax(b, 1)
# if i_activations is not None:
# tf.transpose(tf.transpose(c, perm=[0,2,1]) * i_activations, perm=[0,2,1])
outputs = squash_v1(K.batch_dot(c, u_hat_vecs, [2, 2]))
if i < iterations - 1:
b = b + K.batch_dot(outputs, u_hat_vecs, [2, 3])
poses = outputs
activations = K.sqrt(K.sum(K.square(poses), 2))
return poses, activations
def vec_transformationByConv(poses, input_capsule_dim, input_capsule_num, output_capsule_dim, output_capsule_num):
kernel = _get_weights_wrapper(
name='weights', shape=[1, input_capsule_dim, output_capsule_dim*output_capsule_num], weights_decay_factor=0.0
)
tf.logging.info('poses: {}'.format(poses.get_shape()))
tf.logging.info('kernel: {}'.format(kernel.get_shape()))
u_hat_vecs = keras.backend.conv1d(poses, kernel)
u_hat_vecs = keras.backend.reshape(u_hat_vecs, (-1, input_capsule_num, output_capsule_num, output_capsule_dim))
u_hat_vecs = keras.backend.permute_dimensions(u_hat_vecs, (0, 2, 1, 3))
return u_hat_vecs
def vec_transformationByMat(poses, input_capsule_dim, input_capsule_num, output_capsule_dim, output_capsule_num, shared=True):
inputs_poses_shape = poses.get_shape().as_list()
poses = poses[..., tf.newaxis, :]
poses = tf.tile(
poses, [1, 1, output_capsule_num, 1]
)
if shared:
kernel = _get_weights_wrapper(
name='weights', shape=[1, 1, output_capsule_num, output_capsule_dim, input_capsule_dim], weights_decay_factor=0.0
)
kernel = tf.tile(
kernel, [inputs_poses_shape[0], input_capsule_num, 1, 1, 1]
)
else:
kernel = _get_weights_wrapper(
name='weights', shape=[1, input_capsule_num, output_capsule_num, output_capsule_dim, input_capsule_dim], weights_decay_factor=0.0
)
kernel = tf.tile(
kernel, [inputs_poses_shape[0], 1, 1, 1, 1]
)
tf.logging.info('poses: {}'.format(poses[...,tf.newaxis].get_shape()))
tf.logging.info('kernel: {}'.format(kernel.get_shape()))
u_hat_vecs = tf.squeeze(tf.matmul(kernel, poses[...,tf.newaxis]),axis=-1)
u_hat_vecs = keras.backend.permute_dimensions(u_hat_vecs, (0, 2, 1, 3))
return u_hat_vecs
def capsules_init(inputs, shape, strides, padding, pose_shape, add_bias, name):
with tf.variable_scope(name):
poses = _conv2d_wrapper(
inputs,
shape=shape[0:-1] + [shape[-1] * pose_shape],
strides=strides,
padding=padding,
add_bias=add_bias,
activation_fn=None,
name='pose_stacked'
)
poses_shape = poses.get_shape().as_list()
poses = tf.reshape(
poses, [
-1, poses_shape[1], poses_shape[2], shape[-1], pose_shape
])
beta_a = _get_weights_wrapper(
name='beta_a', shape=[1, shape[-1]]
)
poses = squash_v1(poses, axis=-1)
activations = K.sqrt(K.sum(K.square(poses), axis=-1)) + beta_a
tf.logging.info("prim poses dimension:{}".format(poses.get_shape()))
return poses, activations
def capsule_fc_layer(nets, output_capsule_num, iterations, name):
with tf.variable_scope(name):
poses, i_activations = nets
input_pose_shape = poses.get_shape().as_list()
u_hat_vecs = vec_transformationByConv(
poses,
input_pose_shape[-1], input_pose_shape[1],
input_pose_shape[-1], output_capsule_num,
)
tf.logging.info('votes shape: {}'.format(u_hat_vecs.get_shape()))
beta_a = _get_weights_wrapper(
name='beta_a', shape=[1, output_capsule_num]
)
poses, activations = routing(u_hat_vecs, beta_a, iterations, output_capsule_num, i_activations)
tf.logging.info('capsule fc shape: {}'.format(poses.get_shape()))
return poses, activations
def capsule_flatten(nets):
poses, activations = nets
input_pose_shape = poses.get_shape().as_list()
poses = tf.reshape(poses, [
-1, input_pose_shape[1]*input_pose_shape[2]*input_pose_shape[3], input_pose_shape[-1]])
activations = tf.reshape(activations, [
-1, input_pose_shape[1]*input_pose_shape[2]*input_pose_shape[3]])
tf.logging.info("flatten poses dimension:{}".format(poses.get_shape()))
tf.logging.info("flatten activations dimension:{}".format(activations.get_shape()))
return poses, activations
def capsule_conv_layer(nets, shape, strides, iterations, name):
with tf.variable_scope(name):
poses, i_activations = nets
inputs_poses_shape = poses.get_shape().as_list()
hk_offsets = [
[(h_offset + k_offset) for k_offset in range(0, shape[0])] for h_offset in
range(0, inputs_poses_shape[1] + 1 - shape[0], strides[1])
]
wk_offsets = [
[(w_offset + k_offset) for k_offset in range(0, shape[1])] for w_offset in
range(0, inputs_poses_shape[2] + 1 - shape[1], strides[2])
]
inputs_poses_patches = tf.transpose(
tf.gather(
tf.gather(
poses, hk_offsets, axis=1, name='gather_poses_height_kernel'
), wk_offsets, axis=3, name='gather_poses_width_kernel'
), perm=[0, 1, 3, 2, 4, 5, 6], name='inputs_poses_patches'
)
tf.logging.info('i_poses_patches shape: {}'.format(inputs_poses_patches.get_shape()))
inputs_poses_shape = inputs_poses_patches.get_shape().as_list()
inputs_poses_patches = tf.reshape(inputs_poses_patches, [
-1, shape[0]*shape[1]*shape[2], inputs_poses_shape[-1]
])
i_activations_patches = tf.transpose(
tf.gather(
tf.gather(
i_activations, hk_offsets, axis=1, name='gather_activations_height_kernel'
), wk_offsets, axis=3, name='gather_activations_width_kernel'
), perm=[0, 1, 3, 2, 4, 5], name='inputs_activations_patches'
)
tf.logging.info('i_activations_patches shape: {}'.format(i_activations_patches.get_shape()))
i_activations_patches = tf.reshape(i_activations_patches, [
-1, shape[0]*shape[1]*shape[2]]
)
u_hat_vecs = vec_transformationByConv(
inputs_poses_patches,
inputs_poses_shape[-1], shape[0]*shape[1]*shape[2],
inputs_poses_shape[-1], shape[3],
)
tf.logging.info('capsule conv votes shape: {}'.format(u_hat_vecs.get_shape()))
beta_a = _get_weights_wrapper(
name='beta_a', shape=[1, shape[3]]
)
poses, activations = routing(u_hat_vecs, beta_a, iterations, shape[3], i_activations_patches)
poses = tf.reshape(poses, [
inputs_poses_shape[0], inputs_poses_shape[1],
inputs_poses_shape[2], shape[3],
inputs_poses_shape[-1]]
)
activations = tf.reshape(activations, [
inputs_poses_shape[0],inputs_poses_shape[1],
inputs_poses_shape[2],shape[3]]
)
nets = poses, activations
tf.logging.info("capsule conv poses dimension:{}".format(poses.get_shape()))
tf.logging.info("capsule conv activations dimension:{}".format(activations.get_shape()))
return nets