本节代码定义了一个 TransformerDecoderBlock 类，它是 Transformer 架构中解码器的一个基本模块。这个模块包含了多头自注意力（Multi-Head Attention）、前馈网络（Feed-Forward Network, FFN）和层归一化（Layer Normalization）。

⭐这一节代码理解即可，知道Transformer的关键组成部分：多头自注意力（Multi-Head Attention）、前馈网络（Feed-Forward Network, FFN）和层归一化（Layer Normalization），不同DecoderBlock 有不同的复现方式，本文只给出了自己的实现方式
 

1. 初始化方法

def __init__(self, d_model, dff, dropout):
    super().__init__()

    self.linear1 = nn.Linear(d_model, dff)
    self.activation = nn.GELU()
    self.dropout = nn.Dropout(dropout)
    self.linear2 = nn.Linear(dff, d_model)

    self.norm1 = nn.LayerNorm(d_model)
    self.norm2 = nn.LayerNorm(d_model)
    self.norm3 = nn.LayerNorm(d_model)
    self.dropout1 = nn.Dropout(dropout)
    self.dropout2 = nn.Dropout(dropout)
    self.dropout3 = nn.Dropout(dropout)

    self.mha_block1 = MultiHeadAttention(d_model, num_heads, dropout)
    self.mha_block2 = MultiHeadAttention(d_model, num_heads, dropout)

• d_model：模型的维度，通常是嵌入维度。

• dff：前馈网络的中间层维度。

• dropout：Dropout 的概率。

• num_heads：多头注意力机制中的头数（未在代码中定义，需要传入）。

2. 多头自注意力机制

self.mha_block1 = MultiHeadAttention(d_model, num_heads, dropout)
self.mha_block2 = MultiHeadAttention(d_model, num_heads, dropout)

• MultiHeadAttention 是一个自定义的多头自注意力模块，通常包含查询（Q）、键（K）和值（V）的线性变换，以及多头注意力机制。

• mha_block1 和 mha_block2 分别表示两个多头自注意力模块。

3. 前馈网络

self.linear1 = nn.Linear(d_model, dff)
self.activation = nn.GELU()
self.dropout = nn.Dropout(dropout)
self.linear2 = nn.Linear(dff, d_model)

• 前馈网络由两个线性层组成，中间使用激活函数（如 GELU 或 ReLU）和 Dropout。

• linear1 将输入从 d_model 映射到 dff，linear2 将输出从 dff 映射回 d_model。

4. 层归一化

self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model)

• 层归一化用于稳定训练过程，减少内部协变量偏移。

5. Dropout

self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)

• Dropout 用于防止过拟合，通过随机丢弃一些神经元的输出来增强模型的泛化能力。

6. 前向传播

def forward(self, x, mask=None):
    x = self.norm1(x + self.dropout1(self.mha_block1(x, mask)))
    x = self.norm2(x + self.dropout2(self.mha_block2(x, mask)))
    x = self.norm3(self.linear2(self.dropout(self.activation(self.linear1(x)))))

    return x

• mha_block1 和 mha_block2：两个多头自注意力模块，分别处理输入 x。

• norm1 和 norm2：在每个自注意力模块后应用层归一化。

• linear1 和 linear2：前馈网络的两个线性层，中间使用激活函数和 Dropout。

• norm3：在前馈网络后应用层归一化。

需复现的完整代码（未标红部分为上节提到的多头自注意力机制）

class MultiHeadAttention(nn.Module):
    def __init__(self, d_model, num_heads, dropout):
        super().__init__()
        self.num_heads = num_heads
        self.d_k = d_model // num_heads
        self.q_project = nn.Linear(d_model, d_model)
        self.k_project = nn.Linear(d_model, d_model)
        self.v_project = nn.Linear(d_model, d_model)
        self.o_project = nn.Linear(d_model, d_model)

        self.dropout = nn.Dropout(dropout)

    def forward(self, x, attn_mask=None):
        
        batch_size, seq_len, d_model = x.shape
        Q = self.q_project(x).view(batch_size, seq_len, self.num_heads, self.d_k).transpose(1, 2)
        K = self.q_project(x).view(batch_size, seq_len, self.num_heads, self.d_k).transpose(1, 2)
        V = self.q_project(x).view(batch_size, seq_len, self.num_heads, self.d_k).transpose(1, 2)

        atten_scores = Q @ K.transpose(2, 3) / math.sqrt(self.d_k)

        if attn_mask is not None:
            attn_mask = attn_mask.unsqueeze(1)
            atten_scores = atten_scores.masked_fill(attn_mask == 0, -1e9)

        atten_scores = torch.softmax(atten_scores, dim=-1)
        out = atten_scores @ V
        out = out.transpose(1, 2).contiguous().view(batch_size, seq_len, d_model)
        out = self.o_project(out)
        return self.dropout(out)

class TransformerDecoderBlock(nn.Module):
    def __init__(self, d_model, dff, dropout):
        super().__init__()

        self.linear1 = nn.Linear(d_model, dff)
        self.activation = nn.GELU()
        # self.activation = nn.ReLU()

        self.dropout = nn .Dropout(dropout)
        self.linear2 = nn.Linear(dff, d_model)

        self.norm1 = nn.LayerNorm(d_model)
        self.norm2 = nn.LayerNorm(d_model)
        self.norm3 = nn.LayerNorm(d_model)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)
        self.dropout3 = nn.Dropout(dropout)

        self.mha_block1 = MultiHeadAttention(d_model, num_heads, dropout)
        self.mha_block2 = MultiHeadAttention(d_model, num_heads, dropout)


    def forward(self, x, mask=None):
        x = self.norm1(x + self.dropout1(self.mha_block1(x, mask)))
        x = self.norm2(x + self.dropout2(self.mha_block2(x, mask)))
        x = self.norm3(self.linear2(self.dropout(self.activation(self.linear1(x)))))

        return x