transformer代码笔记----attention.py

import numpy as np
import torch
import torch.nn as nn


class MultiHeadAttention(nn.Module):
    ''' Multi-Head Attention module '''

    def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):
        super().__init__()

        self.n_head = n_head
        self.d_k = d_k
        self.d_v = d_v

        self.w_qs = nn.Linear(d_model, n_head * d_k) #输入512维，输出512维
        self.w_ks = nn.Linear(d_model, n_head * d_k)
        self.w_vs = nn.Linear(d_model, n_head * d_v)
        nn.init.normal_(self.w_qs.weight, mean=0, std=np.sqrt(2.0 / (d_model + d_k))) #权重初始化。mean=0：正态分布的均值；std：标准差
        nn.init.normal_(self.w_ks.weight, mean=0, std=np.sqrt(2.0 / (d_model + d_k)))#np.sqrt():开根号
        nn.init.normal_(self.w_vs.weight, mean=0, std=np.sqrt(2.0 / (d_model + d_v)))

        self.attention = ScaledDotProductAttention(temperature=np.power(d_k, 0.5),
                                                   attn_dropout=dropout)  #np.power(a,b):a的b次方
        self.layer_norm = nn.LayerNorm(d_model) #归一化

        self.fc = nn.Linear(n_head * d_v, d_model) #线性变换，输入512维，输出512维
        nn.init.xavier_normal_(self.fc.weight) #初始化线性变换的权重

        self.dropout = nn.Dropout(dropout) #信息漏失率

    def forward(self, q, k, v, mask=None):
        d_k, d_v, n_head = self.d_k, self.d_v, self.n_head

        sz_b, len_q, _ = q.size()  #获取查询矩阵的大小，长度
        sz_b, len_k, _ = k.size()
        sz_b, len_v, _ = v.size()

        residual = q

        q = self.w_qs(q).view(sz_b, len_q, n_head, d_k) #重塑tensor的shape
        k = self.w_ks(k).view(sz_b, len_k, n_head, d_k)
        v = self.w_vs(v).view(sz_b, len_v, n_head, d_v)

        #.permute（）：改变tensor维度；contiguous方法改变了多维数组在内存中的存储顺序，以便配合view方法使用
        #torch.contiguous()方法首先拷贝了一份张量在内存中的地址，然后将地址按照形状改变后的张量的语义进行排列。
        q = q.permute(2, 0, 1, 3).contiguous().view(-1, len_q, d_k)  # (n*b) x lq x dk
        k = k.permute(2, 0, 1, 3).contiguous().view(-1, len_k, d_k)  # (n*b) x lk x dk
        v = v.permute(2, 0, 1, 3).contiguous().view(-1, len_v, d_v)  # (n*b) x lv x dv

        if mask is not None:
            mask = mask.repeat(n_head, 1, 1)  # (n*b) x .. x ..

        output, attn = self.attention(q, k, v, mask=mask) #进入attention得到输出和注意力关系

        output = output.view(n_head, sz_b, len_q, d_v) #对输出重塑
        output = output.permute(1, 2, 0, 3).contiguous().view(sz_b, len_q, -1)  # b x lq x (n*dv)

        output = self.dropout(self.fc(output)) #输出进行线性变换和dropout
        output = self.layer_norm(output + residual) #残差连接和归一化

        return output, attn


class ScaledDotProductAttention(nn.Module):
    ''' Scaled Dot-Product Attention '''

    def __init__(self, temperature, attn_dropout=0.1):
        super().__init__()
        self.temperature = temperature
        self.dropout = nn.Dropout(attn_dropout)
        self.softmax = nn.Softmax(dim=2)

    def forward(self, q, k, v, mask=None):
        attn = torch.bmm(q, k.transpose(1, 2)) #三维矩阵相乘（qxk）
        attn = attn / self.temperature #qxk除以根号k的维度

        if mask is not None:
            attn = attn.masked_fill(mask.bool(), -np.inf) #-np.inf：负无穷的浮点数

        attn = self.softmax(attn)
        attn = self.dropout(attn)
        output = torch.bmm(attn, v)

        return output, attn