【HuggingFace】Transformers-BertAttention逐行代码解析

        本文基于Hugging Face的2.6.0版本的Transformers包进行解析，不同版本间略有差异，但无伤大雅。

I. Self-attention的Hugging Face实现

(I). 多头自注意力的实现原理

         关于Transformer的讲解不胜其数，这里就不多写了。

        本文主要写一下多头自注意力机制的操作。我们知道，多头自注意力是要分多个head分别进行自注意力操作，然后将每个head的计算结果concate在一起，然后进行全连接层、LayerNorm等操作，最终得到MHA的输出。

        那么，很多人困惑的点在于，如何将输入特征图(或者说输入序列)转换为多头的输入呢？假设batchsize=1，embedding数量为512，embedding维度(即代码中的hidden_size)为128，即输入序列的维度是[1, 512, 128]，head的数量为8个。代码中的操作是将这个[1, 512, 128]直接进行投影变换，投影矩阵的维度均为128×128，得到Q, K, V，如图(2)所示。

        然后通过view函数将其变形为[1, 512, 8, 16]，再通过premute函数交换维度，得到[1, 8, 512, 16]，具体来看就是图(2)中8个不同颜色的512×16的张量，然后就可以进行多头自注意力计算了。算完之后并不需要concate函数进行拼接，因为是一直在进行矩阵运算，并没有将其切开，本身就是拼在一起的。

        其实可以想一想，中前第1-16列是同一种颜色的，17-32列是另外同一种颜色……以此类推，这就是矩阵运算的魅力。

        另外为什么要先用view后用permute来得到[1, 8, 512, 16]这个维度呢？直接view或者reshape不行吗？那绝对是不行的，这个涉及到按行还是按列展开的问题，感兴趣的可以搜一下。

图(1) 输入序列

图(2) 投影后得到Q/K/V

 (II). 代码实现及解析

        BertAttention类的实现：

class BertAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.self = BertSelfAttention(config)  # self-attention
        self.output = BertSelfOutput(config)  # self-attention后的FC层、DropOut、残差连接和LayerNorm
        self.pruned_heads = set()  # set()函数创建一个无序不重复元素集

    def prune_heads(self, heads):
        """剪枝, 一般用不到"""
        if len(heads) == 0:
            return
        mask = torch.ones(self.self.num_attention_heads, self.self.attention_head_size)
        heads = set(heads) - self.pruned_heads  # Convert to set and remove already pruned heads
        for head in heads:
            # Compute how many pruned heads are before the head and move the index accordingly
            head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
            mask[head] = 0
        mask = mask.view(-1).contiguous().eq(1)
        index = torch.arange(len(mask))[mask].long()

        # Prune linear layers
        self.self.query = prune_linear_layer(self.self.query, index)
        self.self.key = prune_linear_layer(self.self.key, index)
        self.self.value = prune_linear_layer(self.self.value, index)
        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)

        # Update hyper params and store pruned heads
        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
        self.pruned_heads = self.pruned_heads.union(heads)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
    ):
        self_outputs = self.self(
            hidden_states, attention_mask, head_mask, encoder_hidden_states, encoder_attention_mask
        )  # 计算self-attention
        attention_output = self.output(self_outputs[0], hidden_states)  # 计算self-attention后的FC层、Norm等

        # 下面这行有个点要注意：self_outputs[1:]和self_outputs[1]不一样, 若self_outputs[1]不存在, 前者返回的是(), 而后者则会报错
        # 另外, 关于元组的定义中, 若要定义一个单元素的元组, 必须加逗号, 即(1,)才为tuple, 而(1)为int
        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
        return outputs

        BertSelfAttention类的实现：

class BertSelfAttention(nn.Module):
    """self-attention计算
    返回值: 若config.output_attentions为默认值(False), 返回的outputs即为self-attention的计算结果
          若config.output_attentions设置为True, outputs里是self-attention的计算结果和attention weight"""
    def __init__(self, config):
        super().__init__()
        # hidden_size为每条embedding的维度, 必须是num_attention_heads的整数倍
        if config.hidden_size % config.num_attention_heads != 0:
            raise ValueError(
                "The hidden size (%d) is not a multiple of the number of attention "
                "heads (%d)" % (config.hidden_size, config.num_attention_heads)
            )
        # config传参教程：https://www.cnblogs.com/douzujun/p/13572694.html or https://www.zhihu.com/question/493902475
        self.output_attentions = config.output_attentions  # 默认为False, 若config.output_attentions设置为True, 会返回注意力权重

        self.num_attention_heads = config.num_attention_heads  # attention head数量
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        # self.all_head_size这个参数看起来多此一举, 但是仔细观察会发现它和config.hidden_size未必是相等的
        # 主要原因是上面采用了剪枝函数。一般不会采用剪枝, 因此两个参数一般是相等的
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size)  # WQ
        self.key = nn.Linear(config.hidden_size, self.all_head_size)  # WK
        self.value = nn.Linear(config.hidden_size, self.all_head_size)  # WV

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)

    def transpose_for_scores(self, x):
        # x的维度为[batchsize, 512, 128], x.size[:-1]的维度为[batchsize, 512]
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)  # new_x_shape的维度为[batchsize, 512, 8, 16]
        x = x.view(*new_x_shape)  # x的新维度为[batchsize, 512, self.num_attention_heads, self.attention_head_size]=[batchsize, 512, 8, 16]
        return x.permute(0, 2, 1, 3)  # 将self.num_attention_heads调整为channel, [batchsize, 8, 512, 16]

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,  # 用于将某些头的注意力计算无效化, 一般用不到
        encoder_hidden_states=None,
        encoder_attention_mask=None,
    ):
        mixed_query_layer = self.query(hidden_states)  # Query

        # If this is instantiated as a cross-attention module, the keys
        # and values come from an encoder; the attention mask needs to be
        # such that the encoder's padding tokens are not attended to.
        # 如果传入了encoder_hidden_states, 说明是作为cross-attention来用, 那么Key、Value和Query的来源就不同了
        if encoder_hidden_states is not None:
            mixed_key_layer = self.key(encoder_hidden_states)
            mixed_value_layer = self.value(encoder_hidden_states)
            attention_mask = encoder_attention_mask
        else:
            mixed_key_layer = self.key(hidden_states)
            mixed_value_layer = self.value(hidden_states)

        # 下面三个的维度均为(batchsize, num_attention_heads, embedding_num(512), attention_head_size)
        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))  # 计算attention weight
        attention_scores = attention_scores / math.sqrt(self.attention_head_size)  # scale
        if attention_mask is not None:
            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
            # 为什么是+而不是×，详见：https://zhuanlan.zhihu.com/p/360988428
            # +和×实现的效果是一样的,无非一个是加一个极小的负数, 一个是乘极小的整数, 都是为了让attention weight在softmax后接近于0
            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.
        # attention_probs不但做了softmax，还用了一次dropout，这是担心attention矩阵太稠密吗?
        # 这里也提到很不寻常, 但是原始Transformer论文就是这么做的
        attention_probs = self.dropout(attention_probs)

        # Mask heads if we want to
        if head_mask is not None:
            attention_probs = attention_probs * head_mask

        context_layer = torch.matmul(attention_probs, value_layer)

        # context_layer即为self-attention的计算结果
        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)  # 恢复input的形状, 即[batchsize, embedding_num(512), hidden_size(128)]

        # 一般self.output_attentions不会设置为True, 默认为False
        # 若设置为True, outputs是一个含两个元素的元组, 第一个元素是self-attention的计算结果, 第二个元素是attention weights
        outputs = (context_layer, attention_probs) if self.output_attentions else (context_layer,)
        return outputs

         BertSelfOutput类的实现：

class BertSelfOutput(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states

参考：BERT源码详解（一）——HuggingFace Transformers最新版本源码解读 - 知乎 (zhihu.com)