-
Notifications
You must be signed in to change notification settings - Fork 1
/
relative_transformer_module.py
195 lines (160 loc) · 7.84 KB
/
relative_transformer_module.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
import tensorflow as tf
import math
import numpy as np
"""
This part is the components of transformer
Thanks for https://github.com/IsaacChanghau/neural_sequence_labeling
and https://github.com/Kyubyong/transformer
and https://github.com/fastnlp/TENER
Original paper: https://arxiv.org/abs/1706.03762
NER version (TENER) paper: https://arxiv.org/abs/1911.04474
"""
def layer_normalize(inputs, epsilon=1e-8):
with tf.variable_scope("layer_norm"):
inputs_shape = inputs.get_shape()
params_shape = inputs_shape[-1:]
mean, variance = tf.nn.moments(inputs, [-1], keep_dims=True)
beta = tf.Variable(tf.zeros(params_shape))
gamma = tf.Variable(tf.ones(params_shape))
normalized = (inputs - mean) / ((variance + epsilon) ** 0.5)
outputs = tf.add(tf.multiply(gamma, normalized), beta)
return outputs
def _shift(BD):
"""
convert:
-3 -2 -1 0 1 2
-3 -2 -1 0 1 2
-3 -2 -1 0 1 2
to:
0 1 2
-1 0 1
-2 -1 0
"""
bsz, n_head, max_len, _ = BD.get_shape().as_list()
zero_pad = tf.zeros(shape=(bsz, n_head, max_len, 1))
BD = tf.reshape(tf.concat([BD, zero_pad], axis=-1), shape=(bsz, n_head, -1, max_len))
BD = tf.reshape(BD[:, :, :-1], shape=(bsz, n_head, max_len, -1))
BD = BD[:, :, :, max_len:]
return BD
def relative_multi_head_attention(x, num_heads, drop_keep_rate=1.0, reuse=None):
# borrowed from: https://github.com/Kyubyong/transformer/blob/master/modules.py
with tf.variable_scope("relative_multi_head_attention", reuse=reuse):
# attention size must consistent with queries（keys）'s -1 dim
batch_size = x.get_shape().as_list()[0]
attention_size = x.get_shape().as_list()[-1]
max_time = x.get_shape().as_list()[-2]
pos_embed = relative_positional_encoding(x, attention_size // num_heads)[0, :, :]
# linear projections, shape=(batch_size, max_time, attention_size)
query = tf.layers.dense(x, attention_size, activation=tf.nn.relu, name="query_project",
kernel_initializer=tf.contrib.layers.xavier_initializer())
# key do not dense in this model
key = x
value = tf.layers.dense(x, attention_size, activation=tf.nn.relu, name="value_project",
kernel_initializer=tf.contrib.layers.xavier_initializer())
# split and concatenation, shape=(batch_size, num_heads, max_time, attention_size / num_heads)
query_ = tf.stack(tf.split(query, num_heads, axis=2), axis=1)
key_ = tf.stack(tf.split(key, num_heads, axis=2), axis=1)
value_ = tf.concat(tf.split(value, num_heads, axis=2), axis=0)
# shape =(num_heads, attention_size / num_heads)
u = tf.get_variable('var_u', shape=[num_heads, attention_size // num_heads],
initializer=tf.glorot_uniform_initializer())
v = tf.get_variable('var_v', shape=[num_heads, attention_size // num_heads],
initializer=tf.glorot_uniform_initializer())
Qu = query_ + u[:, None]
QKuK = tf.einsum('bnqd,bnkd->bnqk', Qu, key_)
vR = tf.einsum('nd,ld->nl', v, pos_embed)[None, :, None]
QR = tf.einsum('bnqd,ld->bnql', query_, pos_embed)
QRvR = QR + vR
QRvR = _shift(QRvR)
attn_outs = QKuK + QRvR
# attn_outs = tf.reshape(attn_outs, shape=(batch_size*num_heads, max_time, max_time))
attn_outs = tf.concat(tf.unstack(attn_outs, axis=1), axis=0)
# key masking
key_masks = tf.sign(tf.abs(tf.reduce_sum(x, axis=-1))) # shape=(batch_size, max_time)
key_masks = tf.tile(key_masks, [num_heads, 1]) # shape=(batch_size * num_heads, max_time)
# shape=(batch_size * num_heads, max_time, max_time)
key_masks = tf.tile(tf.expand_dims(key_masks, 1), [1, tf.shape(x)[1], 1])
paddings = tf.ones_like(attn_outs) * (-2 ** 32 + 1)
# shape=(batch_size, max_time, attention_size)
attn_outs = tf.where(tf.equal(key_masks, 0), paddings, attn_outs)
# activation
attn_outs = tf.nn.softmax(attn_outs)
# query masking
query_masks = tf.sign(tf.abs(tf.reduce_sum(x, axis=-1)))
query_masks = tf.tile(query_masks, [num_heads, 1])
query_masks = tf.tile(tf.expand_dims(query_masks, -1), [1, 1, tf.shape(x)[1]])
attn_outs *= query_masks
# dropout
attn_outs = tf.nn.dropout(attn_outs, drop_keep_rate)
# weighted sum
outputs = tf.matmul(attn_outs, value_)
# restore shape
outputs = tf.concat(tf.split(outputs, num_heads, axis=0), axis=2)
# residual connection
outputs += x
outputs = layer_normalize(outputs)
return outputs
def relative_positional_encoding(inputs, pos_dim,
zero_pad=False, scale=False, reuse=None):
"""Relative Sinusoidal Positional_Encoding.
Args:
inputs: A 3d Tensor with shape of (batch_size, seq_length, embedding_size).
num_units: Output dimensionality
zero_pad: Boolean. If True, all the values of the first row (id = 0) should be constant zero
scale: Boolean. If True, the output will be multiplied by sqrt num_units(check details from paper)
scope: Optional scope for `variable_scope`.
reuse: Boolean, whether to reuse the weights of a previous layer
by the same name.
Returns:
A 'Tensor' rank equals to inputs's
"""
batch_size, seq_length, embedding_size = inputs.get_shape().as_list()
with tf.variable_scope("positional_encoding", reuse=reuse):
position_ind = tf.tile(tf.expand_dims(tf.range(0, seq_length * 2), 0), [batch_size, 1])
# First part of the PE function: sin and cos argument
# *** update：pos range from [0, seq_length) to [-seq_length//2, seq_length//2] ***
position_enc = np.array([
[pos / np.power(10000, (i - i % 2) / embedding_size) for i in range(pos_dim)] # update pos_dim
for pos in range(-seq_length, seq_length)], dtype='float32')
# Second part, apply the cosine to even columns and sin to odds.
position_enc[:, 0::2] = np.sin(position_enc[:, 0::2]) # dim 2i
position_enc[:, 1::2] = np.cos(position_enc[:, 1::2]) # dim 2i+1
# Convert to a tensor
lookup_table = tf.convert_to_tensor(position_enc)
if zero_pad:
lookup_table = tf.concat((tf.zeros(shape=[1, embedding_size]),
lookup_table[1:, :]), 0)
outputs = tf.nn.embedding_lookup(lookup_table, position_ind)
if scale:
outputs = outputs * embedding_size ** 0.5
# pos_emb_outputs = inputs + outputs
return outputs
def feedforward(inputs, num_units, drop_keep_rate=1.0, reuse=None):
'''Point-wise feed forward net.
Args:
inputs: A 3d tensor with shape of [N, T, C].
num_units: A list of two integers.
scope: Optional scope for `variable_scope`.
reuse: Boolean, whether to reuse the weights of a previous layer
by the same name.
Returns:
A 3d tensor with the same shape and dtype as inputs
'''
with tf.variable_scope("multi_head_attention", reuse=reuse):
# Inner layer
params = {"inputs": inputs, "filters": num_units[0], "kernel_size": 1,
"activation": tf.nn.relu, "use_bias": True}
outputs = tf.layers.conv1d(**params)
# dropout
outputs = tf.nn.dropout(outputs, drop_keep_rate)
# Readout layer
params = {"inputs": outputs, "filters": num_units[1], "kernel_size": 1,
"activation": None, "use_bias": True}
outputs = tf.layers.conv1d(**params)
# dropout
outputs = tf.nn.dropout(outputs, drop_keep_rate)
# Residual connection
outputs += inputs
# Normalize
outputs = layer_normalize(outputs)
return outputs