|
|
import torch
|
|
|
from torch import nn
|
|
|
import config
|
|
|
|
|
|
class ScaledDotProductAttention(nn.Module):
|
|
|
""" Scaled Dot-Product Attention """
|
|
|
|
|
|
def __init__(self, scale_factor=16, dropout=config.dropout):
|
|
|
super().__init__()
|
|
|
self.scale_factor = scale_factor
|
|
|
# dropout用于防止过拟合，在前向传播的过程中，让某个神经元的激活值以一定的概率停止工作
|
|
|
self.dropout = nn.Dropout(dropout)
|
|
|
|
|
|
def forward(self, q, k, v, mask=None):
|
|
|
# batch_size: 批量大小
|
|
|
# len_q,len_k,len_v: 序列长度 在这里他们都相等
|
|
|
# n_head: 多头注意力，论文中默认为8
|
|
|
# d_k,d_v: k v 的dim(维度) 默认都是64
|
|
|
# 此时q的shape为(batch_size, n_head, len_q, d_k) (batch_size, 8, len_q, 64)
|
|
|
# 此时k的shape为(batch_size, n_head, len_k, d_k) (batch_size, 8, len_k, 64)
|
|
|
# 此时v的shape为(batch_size, n_head, len_k, d_v) (batch_size, 8, len_k, 64)
|
|
|
# q先除以self.scale_factor，再乘以k的转置(交换最后两个维度(这样才可以进行矩阵相乘))。
|
|
|
# attn的shape为(batch_size, n_head, len_q, len_k)
|
|
|
|
|
|
attn = torch.matmul(q / self.scale_factor, k.transpose(2, 3))
|
|
|
|
|
|
if mask is not None:
|
|
|
|
|
|
"""
|
|
|
用-1e9代替0 -1e9是一个很大的负数 经过softmax之后接近0
|
|
|
# 其一：去除掉各种padding在训练过程中的影响
|
|
|
# 其二，将输入进行遮盖，避免decoder看到后面要预测的东西。（只用在decoder中）
|
|
|
"""
|
|
|
attn = attn.masked_fill(mask.to('cuda:0') == 0, -1e9)
|
|
|
# 先在attn的最后一个维度做softmax 再dropout 得到注意力分数
|
|
|
attn = self.dropout(torch.softmax(attn, dim=-1))
|
|
|
# 最后attn与v矩阵相乘
|
|
|
# output的shape为(batch_size, 8, len_q, 64)
|
|
|
output = torch.matmul(attn, v)
|
|
|
# 返回 output和注意力分数
|
|
|
return output, attn
|
|
|
|
|
|
class MultiHeadAttention(nn.Module):
|
|
|
""" Multi-Head Attention module """
|
|
|
|
|
|
def __init__(self, n_head=config.n_head, d_model=config.input_dim, d_k=config.input_dim//2, d_v=config.hidden_dim//2, dropout=config.dropout):
|
|
|
# 论文中这里的n_head, d_model, d_k, d_v分别默认为8, 512, 64, 64
|
|
|
'''
|
|
|
# q k v先经过不同的线性层，再用ScaledDotProductAttention，最后再经过一个线性层
|
|
|
'''
|
|
|
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, bias=config.bias)
|
|
|
self.w_ks = nn.Linear(d_model, n_head * d_k, bias=config.bias)
|
|
|
self.w_vs = nn.Linear(d_model, n_head * d_v, bias=config.bias)
|
|
|
self.fc = nn.Linear(n_head * d_v, d_model, bias=config.bias)
|
|
|
|
|
|
self.attention = ScaledDotProductAttention(scale_factor=d_k ** 0.5)
|
|
|
|
|
|
self.dropout = nn.Dropout(dropout)
|
|
|
self.layer_norm = nn.LayerNorm(d_model, eps=1e-6) # 默认对最后一个维度初始化
|
|
|
|
|
|
def forward(self, q, k, v, mask=None):
|
|
|
# q, k, v初次输入为含位置信息的嵌入矩阵X，由于要堆叠N次，后面的输入则是上个多头的输出
|
|
|
# q, k, v：batch_size * seq_num * d_model
|
|
|
d_k, d_v, n_head = self.d_k, self.d_v, self.n_head
|
|
|
# len_q, len_k, len_v 为输入的序列长度
|
|
|
# batch_size为batch_size
|
|
|
batch_size, len_q, len_k, len_v = q.size(0), q.size(1), k.size(1), v.size(1)
|
|
|
# 用作残差连接
|
|
|
residual = q
|
|
|
# Pass through the pre-attention projection: b x lq x (n*dv)
|
|
|
# Separate different heads: b x lq x n x dv
|
|
|
# q k v 分别经过一个线性层再改变维度
|
|
|
# 由(batch_size, len_q, n_head*d_k) => (batch_size, len_q, n_head, d_k)
|
|
|
# (batch_size, len_q, 8*64) => (batch_size, len_q, 8, 64)
|
|
|
q = self.layer_norm(q)
|
|
|
k = self.layer_norm(k)
|
|
|
v = self.layer_norm(v)
|
|
|
|
|
|
# 与q,k,v相关矩阵相乘，得到相应的q,k,v向量，d_model=n_head * d_k
|
|
|
q = self.w_qs(q).view(batch_size, len_q, n_head, d_k)
|
|
|
k = self.w_ks(k).view(batch_size, len_k, n_head, d_k)
|
|
|
v = self.w_vs(v).view(batch_size, len_v, n_head, d_v)
|
|
|
|
|
|
# Transpose for attention dot product: b x n x lq x dv
|
|
|
# 交换维度做attention
|
|
|
# 由(batch_size, len_q, n_head, d_k) => (batch_size, n_head, len_q, d_k)
|
|
|
# (batch_size, len_q, 8, 64) => (batch_size, 8, len_q, 64)
|
|
|
q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
|
|
|
|
|
|
|
|
|
# 输出的q为Softmax(QK/d + (1-S)σ)V, attn 为QK/D
|
|
|
q, attn = self.attention(q, k, v, mask=None)
|
|
|
|
|
|
# Transpose to move the head dimension back: b x lq x n x dv
|
|
|
# Combine the last two dimensions to concatenate all the heads together: b x lq x (n*dv)
|
|
|
# (batch_size, 8, len_k, 64) => (batch_size, len_k, 8, 64) => (batch_size, len_k, 512)
|
|
|
q = q.transpose(1, 2).contiguous().view(batch_size, len_q, -1)
|
|
|
|
|
|
# 经过fc和dropout
|
|
|
q = self.dropout(self.fc(q))
|
|
|
|
|
|
# 残差连接 论文中的Add & Norm中的Add
|
|
|
q += residual
|
|
|
# 论文中的Add & Norm中的Norm
|
|
|
q = self.layer_norm(q)
|
|
|
# q的shape为(batch_size, len_q, 512)
|
|
|
# attn的shape为(batch_size, n_head, len_q, len_k)
|
|
|
return q, attn
|
|
|
|
