-
Notifications
You must be signed in to change notification settings - Fork 4
/
keras_capsnet.py
156 lines (111 loc) · 4.65 KB
/
keras_capsnet.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
import numpy as np
import keras.backend as K
from keras.engine.topology import Layer
from keras.layers import Conv2D, Reshape, Lambda
from keras.activations import softmax
def margin_loss(y_true, y_pred,
m_plus=0.9, m_minus=0.1, down_weighting=0.5):
"""
The margin loss defined in the paper.
The default parameters are those used in the paper.
"""
L = y_true*K.square(K.maximum(0.0, m_plus-y_pred)) + down_weighting*(1-y_true)*K.square(K.maximum(0.0, y_pred-m_minus))
return K.mean(K.sum(L, 1))
def squash(vector, epsilon=K.epsilon()):
vector += epsilon
norm = K.sum(K.square(vector), -1, keepdims=True)
scalar_factor = norm / (1 + norm) / K.sqrt(norm)
squashed = scalar_factor * vector
return squashed
def conv2d_caps(input_layer, nb_filters, kernel_size, capsule_size, strides=2):
conv = Conv2D(
filters=nb_filters*capsule_size,
kernel_size=kernel_size,
strides=strides,
padding='valid'
)(input_layer)
conv_shape = conv.shape
nb_capsules= int(conv_shape[1]*conv_shape[2]*nb_filters)
capsules = Reshape(target_shape=(nb_capsules, capsule_size))(conv)
return Lambda(squash, name='primarycap_squash')(capsules)
class CapsuleLength(Layer):
def call(self, inputs, **kwargs):
input_shape = inputs.get_shape().as_list()
x = K.reshape(inputs, shape=[-1, input_shape[1], input_shape[2]])
return K.sqrt(K.sum(K.square(x), -1))
def compute_output_shape(self, input_shape):
return input_shape[:-2]
class Mask(Layer):
"""
A mask layer for decoder to minimize the marginal loss.
"""
def call(self, inputs):
if type(inputs) is list:
assert len(inputs) == 2
inputs, mask = inputs[0], inputs[1]
assert mask.get_shape().as_list()[1] == inputs.get_shape().as_list()[1]
else:
length = K.sqrt(K.sum(K.square(inputs), axis=-1))
mask = K.one_hot(
indices=K.argmax(length, 1),
num_classes=inputs.get_shape().as_list()[1]
)
mask = K.expand_dims(mask, -1)
# [None, nb_classes, 1]
masked = K.batch_flatten(inputs*mask)
return masked
def compute_output_shape(self, input_shape):
if type(input_shape[0]) is tuple:
return tuple([None, input_shape[0][1]])
else:
return tuple([None, input_shape[1]])
class DenseCapsule(Layer):
"""
A fully connected capsule layer which is similar to
the dense layer but replace the neurons to capsules
"""
def __init__(self, capsule_size, nb_capsules, kernel_initializer='glorot_uniform', iterations=5, **kwargs):
super(DenseCapsule, self).__init__(**kwargs)
self.nb_capsules = nb_capsules
self.iterations = iterations
self.capsule_size = capsule_size
self.initializer = kernel_initializer
def build(self, input_shape):
self.prev_shape = input_shape
self.w_ij = self.add_weight(
name='w_ij',
shape=( self.nb_capsules, input_shape[1], self.capsule_size, input_shape[2]),
initializer=self.initializer
)
self.built = True
def batch_dot(self, X, w, axis):
return K.map_fn(lambda x: K.batch_dot(x, w, axis), elems=X)
def _dynamic_routing(self, u_hat, b_ij):
for i in range(self.iterations):
c_ij = softmax(b_ij, axis=1)
s_j = K.batch_dot(c_ij, u_hat, [2, 2])
v_j = squash(s_j)
if i < self.iterations-1:
b_ij += K.batch_dot(v_j, u_hat, [2, 3])
# b_ij = b_ij + K.batch_dot(K.tile(v_j, [1, self.prev_shape[1], 1, 1, 1]), u_hat, [3, 4])
# return K.squeeze(v_j, axis=1)
return v_j
def call(self, inputs):
expanded_input = K.expand_dims(inputs, 1)
print(expanded_input.shape)
expanded_input = K.tile(expanded_input, [1, self.nb_capsules, 1, 1])
print(expanded_input.shape)
u_hat = K.map_fn(lambda x: K.batch_dot(x, self.w_ij, [2, 3]), elems=expanded_input)
b_ij = K.zeros(shape=[K.shape(u_hat)[0], self.nb_capsules, self.prev_shape[1]], dtype=np.float32)
return self._dynamic_routing(u_hat, b_ij)
def compute_output_shape(self, input_shape):
return tuple([None, self.nb_capsules, self.capsule_size, 1])
def test():
from keras.layers import Input
input_layer = Input(shape=(28, 28, 1))
caps1 = conv2d_caps(input_layer, 64, (3, 3), 8)
print(caps1.shape)
caps2 = DenseCapsule(4, 100)(caps1)
print(caps2.shape)
if __name__ == '__main__':
test()