0%

PyTorch——实现自注意力机制(self-attention)

1 原理简述

  Self-Attention Layer 一次检查同一句子中的所有单词的注意力,这使得它成为一个简单的矩阵计算,并且能够在计算单元上并行计算。 此外,Self-Attention Layer 可以使用下面提到的 Multi-Head 架构来拓宽视野,也就是多头注意力机制。Self-Attention Layer 基本结构如下:
在这里插入图片描述

对于每个输入 $\boldsymbol{x}$,首先经过 Embedding 层对每个输入进行编码得到 $\boldsymbol{a_1,a_2,a_3,a_4}$,后将输入特征经过三个全连接层分别得到 Query,Key,Value

  • $\boldsymbol{q^i(Query) = W^q a^i}$;
  • $\boldsymbol{k^i(Key) = W^k a^i}$;
  • $\boldsymbol{v^i(Value) = W^v a^i}$。

$\boldsymbol{W^q, W^k,W^v}$ 由网络训练而来。注意力矩阵是由 Query 和 Key 计算得到,方式由许多种,如点积、缩放点积等。Value 可以看作是信息提取器,将根据单词的注意力提取一个唯一的值,也即某个特征有多少成分被提取出来。下面计算一种注意力矩阵的方式:缩放点积。
在这里插入图片描述
注意力矩阵 $\boldsymbol{A}$ 定义为 Query (giver) 和 Key (receiver) 的内积除以其维度的平方根。 每个单词通过提供 Query 来匹配作为注意力的目标单词的 Key,从而对所有单词产生注意力。为防止注意力分数随维度增大而增大,让注意力矩阵除以向量的维度的开方。 然后对得到的注意力矩阵 $\boldsymbol{A}$ 进行 Softmax 归一化得到 $\boldsymbol{\hat{A}}$,最后将 $\boldsymbol{\hat{A}}$ 乘以 Value 矩阵并相加得到最终的特征 $\boldsymbol{b}$。
在这里插入图片描述

矩阵化如下:
在这里插入图片描述

在上述的 self-attention 中,我们最终只得到一个注意力矩阵,也就是说这个注意力矩阵所关注的信息只偏句子之间的一种关系,但是在时序序列中,往往特征之间不止一种关系,所以我们要提取多个注意力矩阵,这样可以捕获更多的信息,这种注意力机制也就是 多头注意力机制(Multi-Heads)。在实现过程中,我们只需要将原始的 $\boldsymbol{q^i,k^i,v^i}$ 分裂为 $\boldsymbol{n}$ 个就得到 $\boldsymbol{n}$ 头自注意力机制了。
在这里插入图片描述

2 PyTorch 实现

定义 num_attention_heads 为注意力机制的头数,input_size 为输入特征维度,hidden_size 为 $\boldsymbol{q^i,k^i,v^i}$ 的总维度,这样每个头的维度也可以求出,定义为 attention_head_size:

1
2
3
self.num_attention_heads = num_attention_heads
self.attention_head_size = int(hidden_size / num_attention_heads)
self.all_head_size = hidden_size

定义 $\boldsymbol{W^q, W^k,W^v}$,通过全连接网络生成:

1
2
3
self.key_layer = nn.Linear(input_size, hidden_size)
self.query_layer = nn.Linear(input_size, hidden_size)
self.value_layer = nn.Linear(input_size, hidden_size)

使用输入特征乘 $\boldsymbol{W^q, W^k,W^v}$ 得到 Query,Key,Value 矩阵,维度为 $(batch\_size,seq\_len, hidden\_size)$:

1
2
3
key = self.key_layer(x)
query = self.query_layer(x)
value = self.value_layer(x)

求多头注意力机制的 $\boldsymbol{W^q, W^k,W^v}$,头数为 num_attention_heads,并要调换维度,即将 $seq\_len$ 维度与 $num\_attention\_heads$ 维度对换,最终 $\boldsymbol{W^q, W^k,W^v}$ 维度为 $(batch\_size,num\_attention\_heads,seq\_len,attention\_head\_size)$:

1
2
3
4
5
6
7
def trans_to_multiple_heads(self, x):
new_size = x.size()[ : -1] + (self.num_attention_heads, self.attention_head_size)
x = x.view(new_size)
return x.permute(0, 2, 1, 3)
key_heads = self.trans_to_multiple_heads(key)
query_heads = self.trans_to_multiple_heads(query)
value_heads = self.trans_to_multiple_heads(value)

将 $\boldsymbol{Q}$ 和 $\boldsymbol{K}$ 矩阵做点积运算,并进行缩放,得到注意力矩阵的维度为 $(batch\_size,num\_attention\_heads,seq\_len,seq\_len)$:

1
2
attention_scores = torch.matmul(query_heads, key_heads.permute(0, 1, 3, 2))
attention_scores = attention_scores / math.sqrt(self.attention_head_size)

对注意力矩阵进行归一化,归一化的维度为 3,矩阵的维度不发生变化:

1
attention_probs = F.softmax(attention_scores, dim = -1)

将注意力矩阵乘以矩阵 $\boldsymbol{V}$,得到输出特征,维度为 $(batch\_size,num\_attention\_heads,seq\_len,attention\_head\_size)$:

1
context = torch.matmul(attention_probs, value_heads)

将各头的注意力矩阵进行拼接,contiguous() 是将 tensor 的内存变成连续的,否则进行 view 操作时会报错,至于原因可参考:https://blog.csdn.net/kdongyi/article/details/108180250

1
2
3
context = context.permute(0, 2, 1, 3).contiguous()
new_size = context.size()[ : -2] + (self.all_head_size , )
context = context.view(*new_size)

全部代码:

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
import torch
import numpy as np
import torch.nn as nn
import math
import torch.nn.functional as F

class selfAttention(nn.Module) :
def __init__(self, num_attention_heads, input_size, hidden_size):
super(selfAttention, self).__init__()
if hidden_size % num_attention_heads != 0 :
raise ValueError(
"the hidden size %d is not a multiple of the number of attention heads"
"%d" % (hidden_size, num_attention_heads)
)

self.num_attention_heads = num_attention_heads
self.attention_head_size = int(hidden_size / num_attention_heads)
self.all_head_size = hidden_size

self.key_layer = nn.Linear(input_size, hidden_size)
self.query_layer = nn.Linear(input_size, hidden_size)
self.value_layer = nn.Linear(input_size, hidden_size)

def trans_to_multiple_heads(self, x):
new_size = x.size()[ : -1] + (self.num_attention_heads, self.attention_head_size)
x = x.view(new_size)
return x.permute(0, 2, 1, 3)

def forward(self, x):
key = self.key_layer(x)
query = self.query_layer(x)
value = self.value_layer(x)

key_heads = self.trans_to_multiple_heads(key)
query_heads = self.trans_to_multiple_heads(query)
value_heads = self.trans_to_multiple_heads(value)

attention_scores = torch.matmul(query_heads, key_heads.permute(0, 1, 3, 2))
attention_scores = attention_scores / math.sqrt(self.attention_head_size)

attention_probs = F.softmax(attention_scores, dim = -1)

context = torch.matmul(attention_probs, value_heads)
context = context.permute(0, 2, 1, 3).contiguous()
new_size = context.size()[ : -2] + (self.all_head_size , )
context = context.view(*new_size)
return context

测试:

1
2
3
4
features = torch.rand((32, 20, 10))
attention = selfAttention(2, 10, 20)
result = attention.forward(features)
print(result.shape)

结果:

1
torch.Size([32, 20, 20])

参考:
https://blog.csdn.net/beilizhang/article/details/115282604