PyTorch——自注意力（self-attention）机制实现（代码详解）

参考链接

1. https://www.bilibili.com/video/BV1JE411g7XF?p=54
2. https://arxiv.org/abs/1706.03762
3. https://blog.csdn.net/qq_36653505/article/details/83375160

简述自注意力机制（self-attention）

self-attention可以视为一个特征提取层，给定输入特征$a^1,a^2,\cdot \cdot \cdot a^n$，经过self-attention layer，融合每个输入特征，得到新的特征$b^1,b^2,\cdot \cdot \cdot b^n$。具体如下：

设输入特征为$I$，分别将其乘以三个矩阵$W^q$、$W^k$和$W^v$得到$Q$（query）、$K$（key）和$V$（value）三个矩阵；接下来使用矩阵$Q$和$K$的乘积得到注意力矩阵$A$，归一化得到$\hat{A}$；最后，将归一化后的注意力矩阵$\hat{A}$乘上$V$，得到最后的输出特征$O$。

多头自注意力机制（multi-head self-attention）

上述的self-attention中，每个输入特征$a^i$乘上矩阵$W^q$、$W^k$和$W^v$后，分别得到一个向量$q^i$、$k^i$和$v^i$，称为单头自注意力机制。如果将这些向量$q^i$、$k^i$和$v^i$分裂为$n$个就得到$n$头自注意力机制了。公认多头自注意力机制的效果好于单头的，因为前者可以捕获更多维度的信息。示意图如下：

代码实现

设超参数num_attention_heads为自注意力机制的头数，如此，计算出每个头的维度attention_head_size。

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$W^q$、$W^k$和$W^v$三个矩阵。

self.query = nn.Linear(input_size, self.all_head_size)
self.key = nn.Linear(input_size, self.all_head_size)
self.value = nn.Linear(input_size, self.all_head_size)

下面开始逐步计算，需要主要的是计算过程中张量维度的变化。
将输入特征乘以三个矩阵$W^q$、$W^k$和$W^v$，输出的张量此时还没有区分出多个头。维度变化为：input_tensor $\left(batch,n,input\_size\right)$到mixed_query_layer $\left(batch,n,all\_head\_size\right)$

mixed_query_layer = self.query(input_tensor)
mixed_key_layer = self.key(input_tensor)
mixed_value_layer = self.value(input_tensor)

切分为num_attention_heads个头，并变换维度。维度变化为：mixed_query_layer $\left(batch,n,all\_head\_size\right)$到query_layer $\left(batch,num\_attention\_heads,n,attention\_head\_size\right)$

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

query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)

矩阵$Q$和$K$相乘，得到注意力矩阵，并除以向量的维度的开方，防止注意力分数随维度增大而增大。维度变化为：query_layer $\left(batch,num\_attention\_heads,n,attention\_head\_size\right)$到attention_scores $\left(batch,num\_attention\_heads,n,n\right)$

attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))

attention_scores = attention_scores / math.sqrt(self.attention_head_size)

注意力矩阵归一化。维度变化为：attention_scores $\left(batch,num\_attention\_heads,n,n\right)$到attention_probs $\left(batch,num\_attention\_heads,n,n\right)$

attention_probs = nn.Softmax(dim=-1)(attention_scores)

将注意力矩阵乘以矩阵$V$。维度变化为：ttention_probs $\left(batch,num\_attention\_heads,n,n\right)$乘以value_layer $\left(batch,num\_attention\_heads,n,attention\_head\_size\right)$到context_layer $\left(batch,num\_attention\_heads,n,attention\_head\_size\right)$。

context_layer = torch.matmul(attention_probs, value_layer)

变换context_layer维度，为了后面将各头得到的结果拼接。这里的contiguous()是将tensor的内存变成连续的，为后面的view()做准备。维度变化为：context_layer $\left(batch,num\_attention\_heads,n,attention\_head\_size\right)$到context_layer $\left(batch,n,num\_attention\_heads,attention\_head\_size\right)$

context_layer = context_layer.permute(0, 2, 1, 3).contiguous()

将各头的结果拼接起来。维度变化为：context_layer $\left(batch,n,num\_attention\_heads,attention\_head\_size\right)$到context_layer $\left(batch,n,all\_head\_size\right)$

new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
context_layer = context_layer.view(*new_context_layer_shape)

完整代码

class LayerNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-12):
        """Construct a layernorm module in the TF style (epsilon inside the square root).
        """
        super(LayerNorm, self).__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.bias = nn.Parameter(torch.zeros(hidden_size))
        self.variance_epsilon = eps

    def forward(self, x):
        u = x.mean(-1, keepdim=True)
        s = (x - u).pow(2).mean(-1, keepdim=True)
        x = (x - u) / torch.sqrt(s + self.variance_epsilon)
        return self.weight * x + self.bias
        
class SelfAttention(nn.Module):
    def __init__(self, num_attention_heads, input_size, hidden_size, hidden_dropout_prob):
        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.query = nn.Linear(input_size, self.all_head_size)
        self.key = nn.Linear(input_size, self.all_head_size)
        self.value = nn.Linear(input_size, self.all_head_size)

        self.attn_dropout = nn.Dropout(attention_probs_dropout_prob)

        # 做完self-attention 做一个前馈全连接 LayerNorm 输出
        self.dense = nn.Linear(hidden_size, hidden_size)
        self.LayerNorm = LayerNorm(hidden_size, eps=1e-12)
        self.out_dropout = nn.Dropout(hidden_dropout_prob)

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

    def forward(self, input_tensor):
        mixed_query_layer = self.query(input_tensor)
        mixed_key_layer = self.key(input_tensor)
        mixed_value_layer = self.value(input_tensor)

        query_layer = self.transpose_for_scores(mixed_query_layer)
        key_layer = self.transpose_for_scores(mixed_key_layer)
        value_layer = self.transpose_for_scores(mixed_value_layer)

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))

        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
        # [batch_size heads seq_len seq_len] scores
        # [batch_size 1 1 seq_len]

        # attention_scores = attention_scores + attention_mask

        # Normalize the attention scores to probabilities.
        attention_probs = nn.Softmax(dim=-1)(attention_scores)
        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        # Fixme
        attention_probs = self.attn_dropout(attention_probs)
        context_layer = torch.matmul(attention_probs, value_layer)
        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(*new_context_layer_shape)
        hidden_states = self.dense(context_layer)
        hidden_states = self.out_dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)

        return hidden_states