-
Notifications
You must be signed in to change notification settings - Fork 0
/
local_aggregation.py
183 lines (142 loc) · 6.24 KB
/
local_aggregation.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
from keras.layers.core import *
import numpy as np
import tensorflow as tf
from keras import backend as K
from keras.engine.topology import Layer
from keras import activations, regularizers, constraints
class MyPooling(Layer):
'''
Input shape: (batch_size, dim, channel)
Output shape: (batch_size, dim, 1)
'''
def __init__(self, pool_way = 'my_mean', **kwargs):
self.pool_way = getattr(self, pool_way)
super(MyPooling, self).__init__(**kwargs)
def build(self, input_shape):
self.batch_size = input_shape[0]
self.dim = input_shape[1]
self.channel = input_shape[2]
self.filters = input_shape[3]
def call(self, x, mask = None):
return self.pool_way(x)
def compute_output_shape(self, input_shape):
return (input_shape[0], self.dim)
def my_mean(self, x):
return K.mean(x, axis = 3)
def my_max(self, x):
return K.max(x, axis = 3)
class LocalNet(Layer):
'''
Input shape: (batch_size, dim, nb_channel)
Output shape: (batch_size, dim, nb_filter)
'''
def __init__(self, nb_polynomial_order, neighbor_index, neighbor_weight, neighbor_field, my_init = 0.001,
mult = True, learnable = True, fitting_function = 'chebyshev', **kwargs):
if K.backend() != 'tensorflow':
raise Exception("GraphConv Requires Tensorflow Backend.")
self.nb_polynomial_order = nb_polynomial_order
self.neighbor_index = neighbor_index
self.neighbor_weight = neighbor_weight
self.neighbor_field = neighbor_field
self.my_init = my_init
self.mult = mult
self.learnable = learnable
self.fitting_function = getattr(self, fitting_function)
super(LocalNet, self).__init__(**kwargs)
def build(self, input_shape):
self.dim = input_shape[1]
self.nb_channel = input_shape[2]
if self.mult:
self.W_poly_shape = (self.nb_channel, self.nb_polynomial_order)
else:
self.W_poly_shape = (self.nb_polynomial_order,)
if self.learnable:
self.W_poly = K.random_uniform_variable(self.W_poly_shape,-1*self.my_init, 1*self.my_init,
name='{}_W_poly'.format(self.name))
self.trainable_weights = [self.W_poly]
else:
self.W_poly = K.random_uniform_variable(self.W_poly_shape,1,1,
name='{}_W_poly'.format(self.name))
self.trainable_weights = []
self.built = True
def call(self, x, mask = None):
# x shape: (batch_size, dim, channel) => (batch_size, dim, channel, nb_polynomial_order)
y = self.fitting_function(x)
# (batch_size, dim, channel, nb_polynomial_order)*{(nb_channel, nb_polynomial_order) or (nb_polynomial_order, )}
# (batch_size, dim, channel, nb_polynomial_order)
output = tf.multiply(y, self.W_poly)
return self.Agg(output)#tf.transpose(output, (0,1,3,2))
def compute_output_shape(self, input_shape):
return (input_shape[0], self.dim, self.nb_polynomial_order, self.nb_channel)
def Agg(self, x):
# (batch_size, dim, channel, nb_polynomial_order) => (dim, channel, nb_polynomial_order, batch_size)
x = tf.transpose(x, perm = (1,2,3,0))
# (dim, nb_neighbor, channel, nb_polynomial_order, batch_size)
# (dim, channel, nb_polynomial_order, batch_size)
y = K.sum(tf.gather(x, self.neighbor_index), axis = 1)
# (batch_size, channel, nb_polynomial_order, dim)
z = tf.transpose(y, perm = (3,1,2,0))
z = z*self.neighbor_field
# (batch_size, dim, nb_polynomial_order, channel)
z = tf.transpose(z, perm = (0,3,2,1))
return z
def chebyshev(self, x):
# Chebyshev polynomial: [-1,1]
x = tf.expand_dims(x, dim = 3)
if self.nb_polynomial_order > 1:
x0 = tf.ones_like(x, dtype = 'float32')
x1 = 2 * x # x or 2 * x
y = tf.concat(axis = 3, values = [x0,x1])
for _ in range(2, self.nb_polynomial_order):
x2 = 2 * x * x1 - x0
y = tf.concat(axis = 3, values = [y,x2])
x0, x1 = x1, x2
return y
def legendre(self,x):
# Legendre Polynomials : [-1,1]
x = tf.expand_dims(x, dim = 3)
if self.nb_polynomial_order > 1:
x0 = tf.ones_like(x, dtype = 'float32')
x1 = x
y = tf.concat(axis = 3, values = [x0,x1])
for n in range(2, self.nb_polynomial_order):
x2 = 1.0*(2*n-1)/(n) * x * x1 - 1.0*(n-1)/(n)*x0
y = tf.concat(axis = 3, values = [y,x2])
x0, x1 = x1, x2
return y
def laguerre(self,x):
# Laguerre polynomial: [0,+infinity]
x = tf.expand_dims(x, dim = 3)
if self.nb_polynomial_order > 1:
x0 = tf.ones_like(x, dtype = 'float32')
x1 = x0 - x
y = tf.concat(axis = 3, values = [x0,x1])
for n in range(2, self.nb_polynomial_order):
x2 = (2 * n -1) * x1 - x * x1 - (n-1) * (n-1) * x0
y = tf.concat(axis = 3, values = [y,x2])
x0, x1 = x1, x2
return y
def hermite(self, x):
# Hermite polynomial: [-infinity,+infinity]
x = tf.expand_dims(x, dim = 3)
if self.nb_polynomial_order > 1:
x0 = tf.ones_like(x, dtype = 'float32')
x1 = 2 * x
y = tf.concat(axis = 3, values = [x0,x1])
for n in range(2, self.nb_polynomial_order):
x2 = 2 * x * x1 - 2 * (n-1) * x0
y = tf.concat(axis = 3, values = [y,x2])
x0, x1 = x1, x2
return y
def optimum(self, x):
# Optimum polynomial: {1, x, x^{2}, x^{3}, ...,}
x = tf.expand_dims(x, dim = 3)
if self.nb_polynomial_order > 1:
x0 = tf.ones_like(x, dtype = 'float32')
x1 = x
y = tf.concat(axis = 3, values = [x0,x1])
for n in range(2, self.nb_polynomial_order):
x2 = x * x1
y = tf.concat(axis = 3, values = [y,x2])
x0, x1 = x1, x2
return y