-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathlayers.py
153 lines (108 loc) · 5.04 KB
/
layers.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
import math
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras import backend as K
class PositionLayer(tf.keras.layers.Layer):
def __init__(self, embedding_size, **kwargs):
self.embedding_size = embedding_size;
super(PositionLayer, self).__init__(**kwargs)
def build(self, input_shape):
super(PositionLayer, self).build(input_shape)
def call(self, x):
mask = K.expand_dims(K.cast(K.arange(start =0, stop= K.shape(x)[1] +1), 'float32'), axis=-1);
bins = K.expand_dims(K.cast(K.arange(self.embedding_size // 2) * 2, 'float32'), axis=0);
evens = K.dot(mask, 1.0 / K.pow( 10000.0, bins / self.embedding_size));
odds = tf.identity(evens);
evens = K.sin(evens)[1:, :];
odds = K.cos(odds)[1:, :];
pos = K.reshape(K.stack([evens, odds], axis=2), (-1, K.shape(x)[1], self.embedding_size));
return pos;
def compute_output_shape(self, input_shape):
return input_shape + (self.embedding_size,);
class MaskLayerLeft(tf.keras.layers.Layer):
def __init__(self, **kwargs):
super(MaskLayerLeft, self).__init__(**kwargs)
def build(self, input_shape):
super(MaskLayerLeft, self).build(input_shape)
def call(self, x):
length = K.shape(x)[1];
rank = tf.ones(shape=(1, length), dtype='float32');
y = K.expand_dims(x, axis=-1);
mask = K.dot(y, rank);
return tf.transpose(mask, (0,2,1));
def compute_output_shape(self, input_shape):
return input_shape + (input_shape[1]);
class MaskLayerRight(tf.keras.layers.Layer):
def __init__(self, **kwargs):
super(MaskLayerRight, self).__init__(**kwargs)
def build(self, input_shape):
super(MaskLayerRight, self).build(input_shape)
def call(self, x):
right = x[0];
left = x[1];
length = K.shape(right)[1];
rank = tf.ones(shape=(1, length), dtype='float32');
y = K.expand_dims(left, axis=-1);
mask = K.dot(y, rank);
return tf.transpose(mask, (0,2,1));
def compute_output_shape(self, input_shape):
return input_shape + (input_shape[1]);
class MaskLayerTriangular(tf.keras.layers.Layer):
def __init__(self, **kwargs):
super(MaskLayerTriangular, self).__init__(**kwargs)
def build(self, input_shape):
super(MaskLayerTriangular, self).build(input_shape)
def call(self, x):
t = tf.ones(shape= (K.shape(x)[0], K.shape(x)[1], K.shape(x)[1]));
tri = tf.matrix_band_part(t, -1, 0);
rank = tf.ones(shape=(1, K.shape(x)[1]), dtype='float32');
y = K.expand_dims(x, axis=-1);
mask = K.dot(y, rank);
return tri * tf.transpose(mask, (0,2,1));
def compute_output_shape(self, input_shape):
return input_shape + (input_shape[1],);
class LayerNormalization(tf.keras.layers.Layer):
def __init__(self, eps=1e-6, **kwargs):
self.eps = eps
super(LayerNormalization, self).__init__(**kwargs)
def build(self, input_shape):
self.gamma = self.add_weight(name='gamma', shape=input_shape[-1:],
initializer= tf.keras.initializers.Ones(), trainable=True)
self.beta = self.add_weight(name='beta', shape=input_shape[-1:],
initializer= tf.keras.initializers.Zeros(), trainable=True)
super(LayerNormalization, self).build(input_shape)
def call(self, x):
mean = K.mean(x, axis=-1, keepdims=True)
std = K.std(x, axis=-1, keepdims=True)
return self.gamma * (x - mean) / (std + self.eps) + self.beta
def compute_output_shape(self, input_shape):
return input_shape;
class SelfLayer(tf.keras.layers.Layer):
def __init__(self, embedding_size, key_size, **kwargs):
self.embedding_size = embedding_size;
self.key_size = key_size;
self.denom = math.sqrt(embedding_size);
super(SelfLayer, self).__init__(**kwargs);
def build(self, input_shape):
self.K = self.add_weight( shape =(self.embedding_size, self.key_size),
name="K", trainable = True,
initializer = 'glorot_uniform');
self.V = self.add_weight( shape =(self.embedding_size, self.key_size),
name="V", trainable = True,
initializer = 'glorot_uniform');
self.Q = self.add_weight( shape =(self.embedding_size, self.key_size),
name="Q", trainable = True,
initializer = 'glorot_uniform');
super(SelfLayer, self).build(input_shape);
def call(self, inputs):
Q = tf.tensordot(inputs[0], self.Q, axes = [[2], [0]]);
K = tf.tensordot(inputs[1], self.K, axes = [[2], [0]]);
V = tf.tensordot(inputs[2], self.V, axes = [[2], [0]]);
A = tf.keras.backend.batch_dot(Q, tf.transpose(K, (0,2,1)));
A = A / self.denom;
A = tf.exp(A) * inputs[3];
A = A / tf.reshape( tf.reduce_sum(A, axis = 2), (-1, tf.shape(inputs[0])[1] ,1));
A = layers.Dropout(rate = 0.1) (A);
return tf.keras.backend.batch_dot(A, V);
def compute_output_shape(self, input_shape):
return (input_shape[0],input_shape[1], self.key_size);