-
Notifications
You must be signed in to change notification settings - Fork 13
Expand file tree
/
Copy pathself_attention.py
More file actions
154 lines (99 loc) · 5.84 KB
/
self_attention.py
File metadata and controls
154 lines (99 loc) · 5.84 KB
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
try:
import cupy as np
is_cupy_available = True
except:
import numpy as np
is_cupy_available = False
from transformer.layers.base.dense import Dense
from transformer.layers.base.dropout import Dropout
from transformer.activations import Sigmoid, Softmax
class MultiHeadAttention:
"""Multi-HeadAttention"""
def __init__(self, d_model = 512, heads_num = 8, dropout = 0.1, data_type = None):
self.d_model = d_model
self.heads_num = heads_num
self.data_type = data_type
self.d_k, self.d_q, self.d_v = self.d_model // heads_num, self.d_model // heads_num, self.d_model // heads_num #512 / 8 = 64
self.scale = np.sqrt(self.d_k).astype(self.data_type)
self.K_linear = Dense(inputs_num = self.d_model, units_num = self.d_k * heads_num, use_bias = False, data_type = self.data_type) # self.W_K = np.random.randn(self.d_model, self.d_k)
self.Q_linear = Dense(inputs_num = self.d_model, units_num = self.d_q * heads_num, use_bias = False, data_type = self.data_type) # self.W_Q = np.random.randn(self.d_model, self.d_q)
self.V_linear = Dense(inputs_num = self.d_model, units_num = self.d_v * heads_num, use_bias = False, data_type = self.data_type) # self.W_V = np.random.randn(self.d_model, self.d_v)
self.O_linear = Dense(inputs_num = self.d_model, units_num = self.d_v * heads_num, use_bias = True , data_type = self.data_type) # self.W_O = np.random.randn(self.d_model, self.heads_num * self.d_v)
self.activation = Softmax()
self.dropout = Dropout(dropout)
def split_heads_forward(self, x):
batch_size = x.shape[0]
return x.reshape(batch_size, -1, self.heads_num, self.d_k).transpose(0, 2, 1, 3)
def split_heads_backward(self, x):
batch_size = x.shape[0]
#x.transpose(0, 2, 1, 3).reshape(batch_size, self.key_len, self.d_model)
return x.transpose(0, 2, 1, 3).reshape(batch_size, -1, self.heads_num * self.d_k)
def group_heads_forward(self, x):
batch_size = x.shape[0]
return x.transpose(0, 2, 1, 3).reshape(batch_size, -1, self.heads_num * self.d_k)
def group_heads_backward(self, x):
batch_size = x.shape[0]
return x.reshape(batch_size, -1, self.heads_num, self.d_k).transpose(0, 2, 1, 3)
def forward(self, query, key, value, mask, training = True):
self.key_len, self.query_len, self.value_len = key.shape[1], query.shape[1], value.shape[1]
#query = [batch size, query len, hid dim]
#key = [batch size, key len, hid dim]
#value = [batch size, value len, hid dim]
K = self.K_linear.forward(key)
Q = self.Q_linear.forward(query)
V = self.V_linear.forward(value)
# self.K = K.reshape(batch_size, self.heads_num, self.key_len, self.d_k)
# self.Q = Q.reshape(batch_size, self.heads_num, self.query_len, self.d_q)
# self.V = V.reshape(batch_size, self.heads_num, self.value_len, self.d_v)
self.K = self.split_heads_forward(K)
self.Q = self.split_heads_forward(Q)
self.V = self.split_heads_forward(V)
energy = np.matmul(self.Q, self.K.transpose(0, 1, 3, 2)) / self.scale
self.mask = np.asarray(mask)
if self.mask is not None:
self.mask = self.mask[:, np.newaxis, ...]
energy = np.where(self.mask == 0, float('-inf'), energy)#float("-1e20")
attention = self.activation.forward(energy)
self.dropout_attention = self.dropout.forward(attention, training)
output = np.matmul(self.dropout_attention, self.V)
# concat_output = output.reshape(batch_size, self.query_len, self.heads_num * self.d_v) #self.d_model
concat_output = self.group_heads_forward(output)
O = self.O_linear.forward(concat_output)
return O, attention
def backward(self, error):
error = self.O_linear.backward(error)
# error = error.reshape(error.shape[0], self.heads_num, self.query_len, self.d_v)
error = self.group_heads_backward(error)
V_error = np.matmul(self.dropout_attention.transpose(0, 1, 3, 2), error)
# V_error = np.matmul(error.transpose(0, 1, 3, 2), self.dropout_attention) #alter
error = np.matmul(error, self.V.transpose(0, 1, 3, 2))
error = self.dropout.backward(error)
error = self.activation.backward(error)
if self.mask is not None:
error = np.where(self.mask == 0, 0, error)
error /= self.scale
Q_error = np.matmul(error, self.K)
# K_error = np.matmul(error.transpose(0, 1, 3, 2), self.Q)
K_error = np.matmul(self.Q.transpose(0, 1, 3, 2), error) #alter
K_error = K_error.transpose(0, 1, 3, 2)
# V_error = V_error.reshape(V_error.shape[0], self.value_len, self.d_model)
# Q_error = Q_error.reshape(Q_error.shape[0], self.query_len, self.d_model)
# K_error = K_error.reshape(K_error.shape[0], self.key_len, self.d_model)
V_error = self.split_heads_backward(V_error)
Q_error = self.split_heads_backward(Q_error)
K_error = self.split_heads_backward(K_error)
V_error = self.V_linear.backward(V_error)
Q_error = self.Q_linear.backward(Q_error)
K_error = self.K_linear.backward(K_error)
return Q_error, K_error, V_error
def set_optimizer(self, optimizer):
self.K_linear.set_optimizer(optimizer)
self.Q_linear.set_optimizer(optimizer)
self.V_linear.set_optimizer(optimizer)
self.O_linear.set_optimizer(optimizer)
def update_weights(self, layer_num):
layer_num = self.K_linear.update_weights(layer_num)
layer_num = self.Q_linear.update_weights(layer_num)
layer_num = self.V_linear.update_weights(layer_num)
layer_num = self.O_linear.update_weights(layer_num)
return layer_num