多头Attention MultiheadAttention 怎么用？详细解释_多个attention输出后如何使用

import torch
import torch.nn as nn
 
# 定义多头注意力层
embed_dim = 512  # 输入嵌入维度
num_heads = 8  # 注意力头的数量
multihead_attn = nn.MultiheadAttention(embed_dim, num_heads)
 
# 创建一些示例数据
batch_size = 10  # 批次大小
seq_len = 20  # 序列长度
query = torch.rand(seq_len, batch_size, embed_dim)  # 查询张量
key = torch.rand(seq_len, batch_size, embed_dim)  # 键张量
value = torch.rand(seq_len, batch_size, embed_dim)  # 值张量
print(query.shape)
# 计算多头注意力
attn_output, attn_output_weights = multihead_attn(query, key, value)
print("Attention output shape:", attn_output.shape)  # [seq_len, batch_size, embed_dim]
print("Attention weights shape:", attn_output_weights.shape)  # [batch_size, num_heads, seq_len, seq_len]