- 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
- 🍖 原作者:K同学啊 | 接辅导、项目定制
🏡我的环境:
- 语言环境:Python3.11.4
- 编译器:Jupyter Notebook
- torcch版本:2.0.1
import torch
import torch.nn as nnclass MultiHeadAttention(nn.Module):def __init__(self, hid_dim, n_heads, dropout):super().__init__()self.hid_dim = hid_dimself.n_heads = n_heads# hid_dim必须整除assert hid_dim % n_heads == 0# 定义wqself.w_q = nn.Linear(hid_dim, hid_dim)# 定义wkself.w_k = nn.Linear(hid_dim, hid_dim)# 定义wvself.w_v = nn.Linear(hid_dim, hid_dim)self.fc = nn.Linear(hid_dim, hid_dim)self.do = nn.Dropout(dropout)self.scale = torch.sqrt(torch.FloatTensor([hid_dim//n_heads]))def forward(self, query, key, value, mask=None):# Q与KV在句子长度这一个维度上数值可以不一样bsz = query.shape[0]Q = self.w_q(query)K = self.w_k(key)V = self.w_v(value)# 将QKV拆成多组,方案是将向量直接拆开了# (64, 12, 300) -> (64, 12, 6, 50) -> (64, 6, 12, 50)# (64, 10, 300) -> (64, 10, 6, 50) -> (64, 6, 10, 50)# (64, 10, 300) -> (64, 10, 6, 50) -> (64, 6, 10, 50)Q = Q.view(bsz, -1, self.n_heads, self.hid_dim//self.n_heads).permute(0, 2, 1, 3)K = K.view(bsz, -1, self.n_heads, self.hid_dim//self.n_heads).permute(0, 2, 1, 3)V = V.view(bsz, -1, self.n_heads, self.hid_dim//self.n_heads).permute(0, 2, 1, 3)# 第1步,Q x K / scale# (64, 6, 12, 50) x (64, 6, 50, 10) -> (64, 6, 12, 10)attention = torch.matmul(Q, K.permute(0, 1, 3, 2)) / self.scale# 需要mask掉的地方,attention设置的很小很小if mask is not None:attention = attention.masked_fill(mask == 0, -1e10)# 第2步,做softmax 再dropout得到attentionattention = self.do(torch.softmax(attention, dim=-1))# 第3步,attention结果与k相乘,得到多头注意力的结果# (64, 6, 12, 10) x (64, 6, 10, 50) -> (64, 6, 12, 50)x = torch.matmul(attention, V)# 把结果转回去# (64, 6, 12, 50) -> (64, 12, 6, 50)x = x.permute(0, 2, 1, 3).contiguous()# 把结果合并# (64, 12, 6, 50) -> (64, 12, 300)x = x.view(bsz, -1, self.n_heads * (self.hid_dim // self.n_heads))x = self.fc(x)return xquery = torch.rand(64, 12, 300)
key = torch.rand(64, 10, 300)
value = torch.rand(64, 10, 300)
attention = MultiHeadAttention(hid_dim=300, n_heads=6, dropout=0.1)
output = attention(query, key, value)
print(output.shape)
多头注意力机制拓展了模型关注不同位置的能力,赋予Attention层多个“子表示空间”。