人工智能大语言模型注意力机制开发学习日志

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

class SimpleAttention(nn.Module):
    def __init__(self, hidden_size):
        super(SimpleAttention, self).__init__()
        self.hidden_size = hidden_size
        # 定义注意力权重矩阵
        self.attention_weights = nn.Linear(hidden_size, 1, bias=False)
        
    def forward(self, encoder_outputs):
        # encoder_outputs: (batch_size, seq_len, hidden_size)
        # 计算注意力分数
        scores = self.attention_weights(encoder_outputs)  # (batch_size, seq_len, 1)
        scores = scores.squeeze(2)  # (batch_size, seq_len)
        
        # 应用softmax函数得到注意力权重
        attention_weights = F.softmax(scores, dim=1)  # (batch_size, seq_len)
        
        # 使用注意力权重对encoder输出进行加权求和
        context_vector = torch.bmm(
            attention_weights.unsqueeze(1),  # (batch_size, 1, seq_len)
            encoder_outputs  # (batch_size, seq_len, hidden_size)
        )  # (batch_size, 1, hidden_size)
        
        context_vector = context_vector.squeeze(1)  # (batch_size, hidden_size)
        
        return context_vector, attention_weights

# 使用示例
batch_size = 2
seq_len = 5
hidden_size = 128

# 模拟encoder输出
encoder_outputs = torch.randn(batch_size, seq_len, hidden_size)

# 创建注意力机制实例
attention = SimpleAttention(hidden_size)

# 计算注意力
context_vector, attention_weights = attention(encoder_outputs)

print(f"Context vector shape: {context_vector.shape}")
print(f"Attention weights shape: {attention_weights.shape}")

class MultiHeadAttention(nn.Module):
    def __init__(self, d_model, num_heads):
        super(MultiHeadAttention, self).__init__()
        self.d_model = d_model
        self.num_heads = num_heads
        self.d_k = d_model // num_heads
        
        # 定义线性变换矩阵
        self.W_q = nn.Linear(d_model, d_model)
        self.W_k = nn.Linear(d_model, d_model)
        self.W_v = nn.Linear(d_model, d_model)
        self.W_o = nn.Linear(d_model, d_model)
        
    def forward(self, query, key, value, mask=None):
        batch_size = query.size(0)
        
        # 线性变换
        Q = self.W_q(query)  # (batch_size, seq_len, d_model)
        K = self.W_k(key)    # (batch_size, seq_len, d_model)
        V = self.W_v(value)  # (batch_size, seq_len, d_model)
        
        # 分割成多个头
        Q = Q.view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)  # (batch_size, num_heads, seq_len, d_k)
        K = K.view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)  # (batch_size, num_heads, seq_len, d_k)
        V = V.view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)  # (batch_size, num_heads, seq_len, d_k)
        
        # 计算注意力分数
        scores = torch.matmul(Q, K.transpose(-2, -1)) / np.sqrt(self.d_k)  # (batch_size, num_heads, seq_len, seq_len)
        
        # 应用mask（如果有的话）
        if mask is not None:
            scores = scores.masked_fill(mask == 0, -1e9)
        
        # 应用softmax函数
        attention_weights = F.softmax(scores, dim=-1)  # (batch_size, num_heads, seq_len, seq_len)
        
        # 加权求和
        attention_output = torch.matmul(attention_weights, V)  # (batch_size, num_heads, seq_len, d_k)
        
        # 合并多个头
        attention_output = attention_output.transpose(1, 2).contiguous().view(
            batch_size, -1, self.d_model)  # (batch_size, seq_len, d_model)
        
        # 最后的线性变换
        output = self.W_o(attention_output)  # (batch_size, seq_len, d_model)
        
        return output, attention_weights

# 使用示例
batch_size = 2
seq_len = 5
d_model = 128
num_heads = 8

# 模拟输入
query = torch.randn(batch_size, seq_len, d_model)
key = torch.randn(batch_size, seq_len, d_model)
value = torch.randn(batch_size, seq_len, d_model)

# 创建多头注意力机制实例
multi_head_attention = MultiHeadAttention(d_model, num_heads)

# 计算多头注意力
output, attention_weights = multi_head_attention(query, key, value)

print(f"Output shape: {output.shape}")
print(f"Attention weights shape: {attention_weights.shape}")