import torch
import torch.nn as nn
import torch.nn.functional as F
class MultiHeadAttention(nn.Module):
def __init__(self, d_model, num_heads):
super(MultiHeadAttention, self).__init__()
assert d_model % num_heads == 0 # 确保d_model可以被num_heads整除
self.d_model = d_model
self.num_heads = num_heads
self.d_head = d_model // num_heads # 每个头的特征维度
# 定义线性变换
self.W_Q = nn.Linear(d_model, d_model) # (d_model, d_model)
self.W_K = nn.Linear(d_model, d_model) # (d_model, d_model)
self.W_V = nn.Linear(d_model, d_model) # (d_model, d_model)
self.W_O = nn.Linear(d_model, d_model) # (d_model, d_model)
def forward(self, Q, K, V, mask=None):
batch_size = Q.size(0) # (batch_size, seq_len, d_model)
# 将输入通过线性变换得到 Q, K, V
Q = self.W_Q(Q) # (batch_size, seq_len, d_model)
K = self.W_K(K) # (batch_size, seq_len, d_model)
V = self.W_V(V) # (batch_size, seq_len, d_model)
# 将 Q, K, V 按照头数分割
Q = Q.view(batch_size, -1, self.num_heads, self.d_head).transpose(1, 2) # (batch_size, num_heads, seq_len, d_head)
K = K.view(batch_size, -1, self.num_heads, self.d_head).transpose(1, 2) # (batch_size, num_heads, seq_len, d_head)
V = V.view(batch_size, -1, self.num_heads, self.d_head).transpose(1, 2) # (batch_size, num_heads, seq_len, d_head)
# 计算注意力
scores = torch.matmul(Q, K.transpose(-2, -1)) / (self.d_head ** 0.5) # (batch_size, num_heads, seq_len, seq_len)
if mask is not None:
scores = scores.masked_fill(mask == 0, float('-inf')) # (batch_size, num_heads, seq_len, seq_len)
attention_weights = F.softmax(scores, dim=-1) # (batch_size, num_heads, seq_len, seq_len)
attention_output = torch.matmul(attention_weights, V) # (batch_size, num_heads, seq_len, d_head)
# 合并所有头的输出
attention_output = attention_output.transpose(1, 2).contiguous().view(batch_size, -1, self.d_model) # (batch_size, seq_len, d_model)
# 线性变换输出
output = self.W_O(attention_output) # (batch_size, seq_len, d_model)
return output
# 测试 MultiHeadAttention
d_model = 512
num_heads = 8
batch_size = 64
seq_len = 10
mha = MultiHeadAttention(d_model, num_heads)
Q = torch.randn(batch_size, seq_len, d_model) # (64, 10, 512)
K = torch.randn(batch_size, seq_len, d_model) # (64, 10, 512)
V = torch.randn(batch_size, seq_len, d_model) # (64, 10, 512)
output = mha(Q, K, V) # (64, 10, 512)
print(output.shape) # 输出: torch.Size([64, 10, 512])
注释详细解释
nn.Linear 初始化：
self.W_Q, self.W_K, self.W_V, self.W_O 是线性层，将输入的特征维度映射到相同的特征维度（d_model），权重矩阵大小为 (d_model, d_model)。
前向传播：
Q, K, V 的初始维度是 (batch_size, seq_len, d_model)，通过线性变换后仍然是相同的维度。
view 操作将 Q, K, V 从 (batch_size, seq_len, d_model) 变为 (batch_size, num_heads, seq_len, d_head)，然后 transpose 使其变为 (batch_size, num_heads, seq_len, d_head)。
scores 是通过计算查询和键的点积得到的，维度为 (batch_size, num_heads, seq_len, seq_len)。
attention_weights 是应用 softmax 后的注意力权重，维度也为 (batch_size, num_heads, seq_len, seq_len)。
attention_output 是对值矩阵进行加权求和后的结果，维度为 (batch_size, num_heads, seq_len, d_head)。
将所有注意力头的输出合并回 (batch_size, seq_len, d_model)，最后通过线性变换 self.W_O 得到输出，维度为 (batch_size, seq_len, d_model)。
希望这可以帮助你更好地理解代码中的每个步骤和张量的尺寸变化！