Multi-Head Attention是一种注意力机制,是transfomer的核心机制,就是图中黄色框内的部分.
Multi-Head Attention的原理是通过将模型分为多个头,形成多个子空间,让模型关注不同方面的信息。每个头独立进行注意力运算,得到一个注意力权重矩阵。输出的结果再通过线性变换和拼接操作组合在一起。这样可以提高模型的表示能力和泛化性能。
在Multi-Head Attention中,每个头的权重矩阵是随机初始化生成的,并在训练过程中通过梯度下降等优化算法进行更新。通过这种方式,模型可以学习到如何将输入序列的不同部分关联起来,从而捕获更多的上下文信息。
总之,Multi-Head Attention通过将模型分为多个头,形成多个子空间,让模型关注不同方面的信息,提高了模型的表示能力和泛化性能。它的源码实现基于Scaled Dot-Product Attention,通过并行运算和组合输出来实现多头注意力机制。
import torch
from torch import nn
class MultiheadAttention(nn.Module):
def __init__(self,
embed_dim,
num_heads,
att_dropout=0.1,
out_dropout=0.1,
average_attn_weights=True,
use_separate_proj_weight = False,
device=None,
dtype=None):
super(MultiheadAttention, self).__init__()
self.embed_dim = embed_dim
self.num_heads = num_heads
self.att_dropout = nn.Dropout(att_dropout)
self.out_dropout = nn.Dropout(out_dropout)
self.average_attn_weights = average_attn_weights
self.head_dim = embed_dim // num_heads
self.scale = self.head_dim ** 0.5
assert self.embed_dim == self.num_heads * self.head_dim, \
'embed_dim <{}> must be divisible by num_heads <{}>'.format(self.embed_dim, self.num_heads)
self.fuse_heads = nn.Linear(self.embed_dim, self.embed_dim)
factory_kwargs = {'device': device, 'dtype': dtype}
self.use_separate_proj_weight = use_separate_proj_weight # 是否对输入进行线性映射
if not use_separate_proj_weight:
self.in_proj_weight = nn.Parameter(torch.empty((3 * embed_dim, embed_dim), **factory_kwargs))
self.in_proj_bias = nn.Parameter(torch.empty(3 * embed_dim, **factory_kwargs))
self._reset_parameters()
def _reset_parameters(self):
nn.init.xavier_uniform_(self.in_proj_weight)
nn.init.constant_(self.in_proj_bias, 0.)
def forward(self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
identity=None,
query_pos=None,
key_pos=None,
use_separate_proj_weight: bool = False):
'''
Args:
query:
key:
value:
identity:
query_pos:
key_pos:
use_separate_proj_weight: 参考pytorch
Returns:
'''
assert query.dim() == 3 and key.dim() == 3 and value.dim() == 3
assert key.shape == value.shape, f"key shape {key.shape} does not match value shape {value.shape}"
tgt_len, bsz, embed_dim = query.shape # [查询数量 batch数量 特征维度]
src_len, _, _ = key.shape # [被查询数量,_,_]
# 默认和query进行shortcut(要在位置编码前,因为output为输出特征,特征和原特征shortcut,下一层再重新加位置编码,否则不就重了)
if identity is None:
identity = query.clone()
# 位置编码
if query_pos is not None:
query = query + query_pos
if key_pos is not None:
key = key + key_pos
# 是否需要对输入进行映射,mmcv中 q=k=v,那么就需要此处进行映射
if not self.use_separate_proj_weight:
assert self.in_proj_weight is not None, "use_separate_proj_weight is False but in_proj_weight is None"
query, key, value = nn.functional._in_projection_packed(query, key, value, self.in_proj_weight, self.in_proj_bias)
# 特征划分为self.num_heads 份 [tgt,b,embed_dim] -> [b,n_h, tgt, d_h]
# [n,b,n_h*d_h] -> [b,n_h,n,d_h] 主要是target和source之前的特征匹配和提取, batch和n_h维度不处理
query = query.contiguous().view(tgt_len, bsz, self.num_heads, self.head_dim).permute(1, 2, 0, 3)
key = key.contiguous().view(src_len, bsz, self.num_heads, self.head_dim).permute(1, 2, 0, 3)
value = value.contiguous().view(src_len, bsz, self.num_heads, self.head_dim).permute(1, 2, 0, 3)
# [b,n_h,tgt_len,src_len]
# Scaled Dot-Product Attention
attention = query @ key.transpose(-2, -1)
attention /= self.scale # 参考: https://blog.csdn.net/zwhdldz/article/details/135462127
attention = torch.softmax(attention, dim=-1) # 行概率矩阵
attention = self.att_dropout(input=attention) # 正则化方法 DropKey,用于缓解 Vision Transformer 中的过拟合问题
# [b,n_h,tgt_len,d_h] = [b,n_h,tgt_len,src_len] * [b,n_h,src_len,d_h]
output = attention @ value
# [b,n_h,tgt_len,d_h] -> [b,tgt_len,embed_dim]
output = output.permute(0, 2, 1, 3).contiguous().view(tgt_len, bsz, embed_dim)
# 头之间通过全连接融合一下
output = self.fuse_heads(output)
output = self.out_dropout(output)
# shortcut
output = output + identity
# 多头head求平均
if self.average_attn_weights:
attention = attention.sum(dim=1) / self.num_heads
# [tgt_len,b,embed_dim],[b,tgt_len,src_len]
return output, attention
if __name__ == '__main__':
query = torch.rand(size=(10, 2, 64))
key = torch.rand(size=(5, 2, 64))
value = torch.rand(size=(5, 2, 64))
query_pos = torch.rand(size=(10, 2, 64))
key_pos = torch.rand(size=(5, 2, 64))
att = MultiheadAttention(64, 4)
# 返回特征采样结果和attention矩阵
output = att(query=query, key=key, value=value,query_pos=query_pos,key_pos=key_pos)
pass
在实现中,参考pytorch我在内部加输入映射,具体作用参考下一篇博客。