-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmylayers.py
159 lines (149 loc) · 8.39 KB
/
mylayers.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
import tensorflow as tf
from utils import Misc
import numpy as np
class RCEncoding(tf.keras.layers.Layer):
# att_mode: additive v^T tanh(Wx + U q)
# att_mode: dot x^T q
# att_mode: scale_dot x^T q / sqrt(d)
# att_mode: bilinear x^T W q
# att_mode: euclidean exp - 0.5|x - q|
# att_mode: kd
# rnn_mode: res x - f(x)
# rnn_mode: rnn W x - U f(x) = [W, U] * [x; -f(x)]
# rnn_mode: none x
def __init__(self, num_codewords, att_mode, rnn_mode, **kwargs):
self.num_codewords = num_codewords
self.T = 0.9
self.prob = 0.5
self.att_mode = att_mode
self.rnn_mode = rnn_mode
super(RCEncoding, self).__init__(**kwargs)
def build(self, input_shape):
self.dim = input_shape.as_list()[-1]
self.codebooks = [None] * len(self.num_codewords)
if self.att_mode == 'additive':
self.W_att_c = [None] * len(self.num_codewords)
self.W_att_x = [None] * len(self.num_codewords)
self.v_att = [None] * len(self.num_codewords)
elif self.att_mode == 'bilinear':
self.W_att = [None] * len(self.num_codewords)
elif self.att_mode == 'kd':
self.W_att = [None] * len(self.num_codewords)
for i in range(len(self.num_codewords)):
self.codebooks[i] = self.add_weight(name='codebook_%d' % i,
shape=(self.num_codewords[i], self.dim),
initializer=tf.keras.initializers.glorot_normal(),
trainable=True)
if self.att_mode == 'additive':
self.W_att_c[i] = self.add_weight(name='weight_att_additive_c_%d' % i,
shape=(self.dim, self.dim),
initializer=tf.keras.initializers.glorot_normal(),
trainable=True)
self.W_att_x[i] = self.add_weight(name='weight_att_additive_x_%d' % i,
shape=(self.dim, self.dim),
initializer=tf.keras.initializers.glorot_normal(),
trainable=True)
self.v_att[i] = self.add_weight(name='v_att_additive_%d' % i,
shape=(self.dim, ),
initializer=tf.keras.initializers.glorot_normal(),
trainable=True)
elif self.att_mode == 'bilinear':
self.W_att[i] = self.add_weight(name='weight_att_bilinear_%d' % i,
shape=(self.dim, self.dim),
initializer=tf.keras.initializers.glorot_normal(),
trainable=True)
elif self.att_mode == 'kd':
self.W_att[i] = self.add_weight(name='weight_att_kd_%d' % i,
shape=(self.dim, self.num_codewords[i]),
initializer=tf.keras.initializers.glorot_normal(),
trainable=True)
if len(self.num_codewords) > 1:
if self.rnn_mode == 'rnn':
self.W_rnn = self.add_weight(name='weight_rnn',
shape=(2*self.dim, self.dim),
initializer=tf.keras.initializers.glorot_normal(),
trainable=True)
elif self.rnn_mode == 'res':
self.W_rnn = self.add_weight(name='weight_rnn',
shape=(self.dim, self.dim),
initializer=tf.keras.initializers.glorot_normal(),
trainable=True)
super(RCEncoding, self).build(input_shape)
def call(self, inputs, **kwargs):
output = [None] * len(self.num_codewords)
indexes = [None] * len(self.num_codewords)
weights = [None] * len(self.num_codewords)
x = inputs
for i in range(len(self.num_codewords)):
if i>0:
if self.rnn_mode == 'rnn':
x = tf.tensordot(tf.concat((x, output[i-1]), -1), self.W_rnn, axes=[-1, 0])
elif self.rnn_mode == 'res':
#x = x - output[i-1]
x = x - tf.tensordot(output[i-1], self.W_rnn, axes=[-1, 0])
if self.att_mode == 'additive':
x_ = tf.tensordot(x, self.W_att_x[i], axes=[-1, 0])
c_ = tf.tensordot(self.codebooks[i], self.W_att_c[i], axes=[-1, 0])
xc = tf.tanh(tf.expand_dims(x_, -2) + c_)
att_logit = tf.tensordot(xc, self.v_att[i], axes=[-1, 0])
elif self.att_mode == 'dot':
att_logit = tf.tensordot(x, self.codebooks[i], axes=[-1, -1])
elif self.att_mode == 'scale_dot':
att_logit = tf.tensordot(x, self.codebooks[i], axes=[-1, -1]) / tf.sqrt(tf.cast(self.dim,dtype=tf.float32))
elif self.att_mode == 'bilinear':
x_ = tf.tensordot(x, self.W_att[i], axes=[-1, 0])
c = tf.tensordot(self.codebooks[i], self.W_att[i], axes=[-1, 0])
att_logit = tf.tensordot(x_, c, axes=[-1, -1])
elif self.att_mode == 'euclidean':
x_ = tf.expand_dims(x, -2)
dist = tf.reduce_sum(tf.squared_difference(x_, self.codebooks[i]), axis=-1, keepdims=False)
att_logit = tf.exp(-dist / 2)
elif self.att_mode == 'kd':
att_logit = tf.tensordot(x, self.W_att[i], axes=[-1, 0])
else:
raise ValueError('Not supported attention mode')
weight_soft = Misc.gumbel_softmax(att_logit, self.T, hard=False)
weight_hard = Misc.gumbel_softmax(att_logit, self.T, hard=True)
indexes[i] = tf.argmax(att_logit, -1)
weights[i] = weight_soft
weight = tf.stop_gradient(weight_hard - weight_soft) + weight_soft
#weight = tf.keras.backend.in_train_phase(weight_soft, weight_hard)
#weight = tf.keras.backend.in_train_phase(weight_, weight_hard)
output[i] = tf.tensordot(weight, self.codebooks[i], axes=[-1, 0])
#train_output = tf.nn.dropout(tf.add_n(output), rate=1-self.prob) + tf.nn.dropout(inputs, rate=self.prob)
#test_output = tf.add_n(output)
#return tf.keras.backend.in_train_phase(train_output, test_output), tf.concat(indexes, -1)
return tf.add_n(output), tf.concat(indexes, -1)
def compute_output_shape(self, input_shape):
output_shape = []
output_shape.extend(input_shape[:-1])
output_shape.append(len(self.num_codewords))
return input_shape, output_shape
class EuclideanDistillation(tf.keras.layers.Layer):
def __init__(self, coef=1, **kwargs):
self.coef = coef
super(EuclideanDistillation, self).__init__(**kwargs)
def call(self, inputs):
x, x_ = inputs
self.add_loss(self.coef*tf.reduce_mean((x - x_)**2))
return x_
class DistilledDot(tf.keras.layers.Layer):
def __init__(self, ed_coef=1, id_coef=0, **kwargs):
self.ed_coef = ed_coef
self.id_coef = id_coef
self.momentum = 0.99
super(DistilledDot, self).__init__(**kwargs)
def build(self, input_shape):
dim = input_shape[0].as_list()[-1]
self.moving_user = self.add_weight(name='moving_user', shape=(dim, dim), trainable=False, initializer='zeros')
self.moving_user_ = self.add_weight(name='moving_user_', shape=(dim, dim), trainable=False, initializer='zeros')
def call(self, inputs):
user, user_, item, item_ = inputs
if self.ed_coef>0:
self.add_loss(self.ed_coef * tf.reduce_mean((item - item_)**2))
if self.id_coef>0:
x = tf.tensordot(user, user, axes=[[0,1], [0,1]])
x_ = tf.tensordot(user_, user_, axes=[[0,1], [0,1]])
self.add_update(tf.keras.backend.moving_average_update(self.moving_user, x, self.momentum))
self.add_update(tf.keras.backend.moving_average_update(self.moving_user_, x_, self.momentum))
return tf.tensordot(user_, item_, axes=-1)