Multi-Head Self-Attention

前言

之前手搓了Self-Attention，这次来搓一个最基础配置的Multi-Head Self-Attention，关于Self-Attention的构造思路可以看我之前的博客：手搓Self-Attention。

步骤

同样根据大佬的博客，我们分析一下，在图例中：

，，，即

我们之前已经完成了Z的计算，所以只需要将各个Z拼接在一起，然后做线性变换得到最终的输出即可。

实现

1. Self-Attention

class SelfAttention(nn.Module):
    def __init__(self, d_m, d_k) -> None:
        super(SelfAttention, self).__init__()
        self.d_m = d_m
        self.d_k = d_k
        self.Wq = nn.Linear(in_features=self.d_m, out_features=self.d_k)
        self.Wk = nn.Linear(in_features=self.d_m, out_features=self.d_k)
        self.Wv = nn.Linear(in_features=self.d_m, out_features=self.d_k)

    def forward(self, x):
        q,k,v = self.Wq(x), self.Wk(x),self.Wv(x)
        score = pt.einsum('nci, ncj -> nc',q,k)
        score /= pt.sqrt(pt.tensor(self.d_k))
        score = nn.functional.softmax(score, dim=-1)
        out = v * score[:,:,None]
        return out

2. Multi-Head Self-Attention

这里我们遵从nn.MultiheadAttention，模型的维度必须整除头数（图示中头数是8大于嵌入维度是4，有些奇怪，是哪里出了问题，大佬可以评论一下）。

class MultiHeadSelfAttention(nn.Module):
    def __init__(self, embed_dim, num_heads) -> None:
        super(MultiHeadSelfAttention, self).__init__()
        assert embed_dim % num_heads == 0, 'embed_dim must be divided by num_heads'
        self.heads = nn.ModuleList([SelfAttention(d_m=embed_dim, d_k=embed_dim//num_heads) 
                                    for _ in range(num_heads)])
        self.Wo = nn.Linear(in_features=embed_dim, out_features=embed_dim)
    
    def forward(self, x):
        Z = pt.cat([head(x) for head in self.heads], dim=-1)
        return self.Wo(Z)


''' example '''
a = pt.randn(2, 3, 10)
m = MultiHeadSelfAttention(10, 5)
b = m(a)
b

效果

左边是单头，右边是双头，效果略好于单头。

结尾

有问题欢迎在评论区讨论！