多头自注意力机制的代码实现

• transformer的整体结构：
  

1、自注意力机制

• 自注意力机制如下：
  
• 计算过程：
  
• 代码如下：

class ScaledDotProductAttention(nn.Module):
    def __init__(self, embed_dim, key_size, value_size):
        super().__init__()
		self.W_q = nn.Linear(embed_dim, key_size, bias=False)
        self.W_k = nn.Linear(embed_dim, key_size, bias=False)
        self.W_v = nn.Linear(embed_dim, value_size, bias=False)

    def forward(self, x, attn_mask=None):
        """
        Args:
            X: shape: (N, L, embed_dim), input sequence, 
            是经过input embedding后的输入序列，L个embed_dim维度的嵌入向量
            attn_mask: (N, L, L)，用于对注意力矩阵(L, L)进行mask
        输出：shape：(N, L, embed_dim)
        """
        query = self.W_q(x)  # (N, L, key_size)
        key = self.W_k(x)  # (N, L, key_size)
        value = self.W_v(x)  # (N, L, value_size)
        scores = torch.matmul(query, key.transpose(1, 2)) / math.sqrt(query.size(2))
        if attn_mask is not None:
            scores = scores.masked_fill(attn_mask, 0)
        attn_weights = F.softmax(scores, dim=-1)	# dim为-1表示，对每个嵌入向量与其他所有向量的注意力权重，进行softmax，以使每一行的和为1
        return torch.matmul(attn_weights, value)

2、多头注意力机制

• 结构如下：
  
• 计算过程如下：

class MultiHeadSelfAttention(nn.Module):
    def __init__(self, embed_dim, num_heads, key_size, value_size, bias=False):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.q_head_dim = key_size // num_heads
		self.k_head_dim = key_size // num_heads
		self.v_head_dim = value_size // num_heads
        
		self.W_q = nn.Linear(embed_dim, key_size, bias=bias)
        self.W_k = nn.Linear(embed_dim, key_size, bias=bias)
        self.W_v = nn.Linear(embed_dim, value_size, bias=bias)        

        self.q_proj = nn.Linear(key_size, key_size, bias=bias)
        self.k_proj = nn.Linear(key_size, key_size, bias=bias)
        self.v_proj = nn.Linear(value_size, value_size, bias=bias)
        self.out_proj = nn.Linear(value_size, embed_dim, bias=bias)

    def forward(self, x):
        """
        Args:
            X: shape: (N, L, embed_dim), input sequence, 
            是经过input embedding后的输入序列，L个embed_dim维度的嵌入向量

        Returns:
            output: (N, L, embed_dim)
        """
        query = self.W_q(x)  # (N, L, key_size)
        key = self.W_k(x)  # (N, L, key_size)
        value = self.W_v(x)  # (N, L, value_size)
        q, k, v = self.q_proj(query), self.k_proj(key), self.v_proj(value)
        N, L, value_size = v.size()

        q = q.reshape(N, L, self.num_heads, self.q_head_dim).transpose(1, 2)
        k = k.reshape(N, L, self.num_heads, self.k_head_dim).transpose(1, 2)
        v = v.reshape(N, L, self.num_heads, self.v_head_dim).transpose(1, 2)

		att = torch.matmul(q, k.transpose(-1, -2)) / math.sqrt(k.size(-1))
		att = F.softmax(att, dim=-1)
		output = torch.matmul(att, v)
		output = output.transpose(1, 2).reshape(N, L, value_size)
		output = self.out_proj(output)
		
        return output

参考：
https://blog.csdn.net/weixin_43334693/article/details/130313746