英译法案例 - 基于 GRU 有 Attention 的解码器思路分析

import torch
import torch.nn as nn
import torch.nn.functional as F


class DecoderGRUWithAttention(nn.Module):
    """基于 GRU 的有 Attention 解码器"""
    
    def __init__(self, hidden_size, output_size, dropout_p=0.1, max_length=10):
        """
        初始化函数
        
        Args:
            hidden_size: 隐藏层维度（256）
            output_size: 输出维度（4345，法语词汇表大小）
            dropout_p: 随机失活概率
            max_length: 句子最大长度
        """
        super(DecoderGRUWithAttention, self).__init__()
        self.hidden_size = hidden_size
        self.output_size = output_size
        self.dropout_p = dropout_p
        self.max_length = max_length
        
        # 1. 词嵌入层
        self.embedding = nn.Embedding(output_size, hidden_size)
        
        # 2. 随机失活层
        self.dropout = nn.Dropout(dropout_p)
        
        # 3. 注意力权重计算层（Q 和 K 拼接后 512 → 10）
        self.attn = nn.Linear(hidden_size * 2, max_length)
        
        # 4. 注意力应用后的线性层（attn_applied + Q 拼接后 512 → 256）
        self.attn_combine = nn.Linear(hidden_size * 2, hidden_size)
        
        # 5. GRU 层
        self.gru = nn.GRU(hidden_size, hidden_size)
        
        # 6. 线性层
        self.out = nn.Linear(hidden_size, output_size)
        
        # 7. LogSoftmax
        self.softmax = nn.LogSoftmax(dim=1)
    
    def forward(self, input, prev_hidden, encoder_outputs):
        """
        前向传播
        
        Args:
            input: 当前时间步的输入词索引 [1, 1]
            prev_hidden: 上一个时间步的隐藏状态 [1, 1, 256]
            encoder_outputs: 编码器所有时间步的输出 [1, 10, 256]
        
        Returns:
            output: 概率分布 [1, 4345]
            hidden: 本次隐藏状态 [1, 1, 256]
            attn_weights: 注意力权重分布 [1, 1, 10]
        """
        # 1. 词嵌入：[1, 1] → [1, 1, 256]
        embedded = self.embedding(input).view(1, 1, -1)
        embedded = self.dropout(embedded)
        
        # 2. 计算注意力权重
        # Q 和 K 拼接：[1, 256] + [1, 256] → [1, 512]
        attn_weights = F.softmax(
            self.attn(torch.cat((embedded[0], prev_hidden[0]), 1)),  # [1, 512] → [1, 10]
            dim=1
        )
        
        # 3. 注意力应用
        # attn_weights: [1, 1, 10], encoder_outputs: [1, 10, 256]
        attn_applied = torch.bmm(attn_weights.unsqueeze(0), encoder_outputs.unsqueeze(0))  # [1, 1, 256]
        
        # 4. 拼接 attn_applied 和 Q
        # [1, 256] + [1, 256] → [1, 512]
        output = torch.cat((attn_applied[0], embedded[0]), 1)  # [1, 512]
        
        # 5. 线性层：[1, 512] → [1, 256]
        output = self.attn_combine(output).unsqueeze(0)  # [1, 1, 256]
        output = F.relu(output)
        
        # 6. GRU：[1, 1, 256] → [1, 1, 256], [1, 1, 256]
        output, hidden = self.gru(output, prev_hidden)
        
        # 7. 线性层 + LogSoftmax：[1, 256] → [1, 4345]
        output = self.softmax(self.out(output[0]))
        
        return output, hidden, attn_weights


# 测试
if __name__ == "__main__":
    print("=== 英译法案例 - 基于 GRU 有 Attention 的解码器 ===")
    
    hidden_size = 256
    output_size = 4345
    max_length = 10
    
    decoder = DecoderGRUWithAttention(hidden_size, output_size, dropout_p=0.1, max_length=max_length)
    print(f"解码器:\n{decoder}")
    
    # 创建输入
    input_tensor = torch.tensor([[1]])  # [1, 1]
    prev_hidden = torch.zeros(1, 1, hidden_size)  # [1, 1, 256]
    encoder_outputs = torch.randn(1, max_length, hidden_size)  # [1, 10, 256]
    
    # 前向传播
    output, hidden, attn_weights = decoder(input_tensor, prev_hidden, encoder_outputs)
    
    print(f"\n输出形状: {output.shape}")  # [1, 4345]
    print(f"隐藏状态形状: {hidden.shape}")  # [1, 1, 256]
    print(f"注意力权重形状: {attn_weights.shape}")  # [1, 10]