PyTorch 2.8 实战：从零复现经典论文《Attention Is All You Need》-编程实验室

PyTorch 2.8 实战：从零复现经典论文《Attention Is All You Need》

1. 引言：Transformer为何如此重要

2017年，一篇名为《Attention Is All You Need》的论文彻底改变了人工智能领域的发展轨迹。这篇论文提出的Transformer架构，如今已成为自然语言处理乃至整个AI领域的基石。从ChatGPT到Stable Diffusion，几乎所有现代AI系统背后都有Transformer的身影。

本文将带你用PyTorch 2.8从零开始复现这篇经典论文。不同于简单调用现成的Transformer模块，我们会逐行实现论文中的关键组件，包括：

多头注意力机制（Multi-Head Attention）
位置编码（Positional Encoding）
前馈网络（Feed Forward Network）
完整的编码器-解码器架构

通过这个实战项目，你不仅能深入理解Transformer的工作原理，还能掌握PyTorch 2.8的最新特性在实际项目中的应用技巧。

2. 环境准备与数据加载

2.1 快速搭建开发环境

首先确保你已安装Python 3.8+和PyTorch 2.8。推荐使用conda创建虚拟环境：

conda create -n transformer python=3.9 conda activate transformer pip install torch==2.8.0 torchtext==0.16.0

2.2 准备机器翻译数据集

我们将使用论文中提到的IWSLT 2016德英翻译数据集。这个数据集规模适中，非常适合教学目的：

from torchtext.datasets import IWSLT2016 from torchtext.data.utils import get_tokenizer # 加载数据集和分词器 train_iter = IWSLT2016(split='train', language_pair=('de', 'en')) tokenizer_de = get_tokenizer('spacy', language='de_core_news_sm') tokenizer_en = get_tokenizer('spacy', language='en_core_web_sm')

3. 实现Transformer核心组件

3.1 位置编码：让模型理解顺序信息

Transformer没有循环结构，需要通过位置编码注入序列顺序信息。以下是论文中的正弦/余弦位置编码实现：

import torch import math class PositionalEncoding(torch.nn.Module): def __init__(self, d_model, max_len=5000): super().__init__() pe = torch.zeros(max_len, d_model) position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1) div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)) pe[:, 0::2] = torch.sin(position * div_term) pe[:, 1::2] = torch.cos(position * div_term) pe = pe.unsqueeze(0) self.register_buffer('pe', pe) def forward(self, x): return x + self.pe[:, :x.size(1)]

3.2 多头注意力机制：Transformer的灵魂

这是Transformer最具创新性的部分，让我们分解实现：

class MultiHeadAttention(torch.nn.Module): def __init__(self, d_model, num_heads): super().__init__() assert d_model % num_heads == 0 self.d_k = d_model // num_heads self.num_heads = num_heads self.q_linear = torch.nn.Linear(d_model, d_model) self.v_linear = torch.nn.Linear(d_model, d_model) self.k_linear = torch.nn.Linear(d_model, d_model) self.out = torch.nn.Linear(d_model, d_model) def forward(self, q, k, v, mask=None): batch_size = q.size(0) # 线性变换并分头 q = self.q_linear(q).view(batch_size, -1, self.num_heads, self.d_k) k = self.k_linear(k).view(batch_size, -1, self.num_heads, self.d_k) v = self.v_linear(v).view(batch_size, -1, self.num_heads, self.d_k) # 转置以获得形状 (batch_size, num_heads, seq_len, d_k) q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2) # 计算注意力分数 scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k) if mask is not None: scores = scores.masked_fill(mask == 0, -1e9) scores = torch.nn.functional.softmax(scores, dim=-1) # 应用注意力权重到V output = torch.matmul(scores, v) # 合并多头 output = output.transpose(1, 2).contiguous().view(batch_size, -1, self.d_model) return self.out(output)

4. 构建完整Transformer模型

4.1 编码器层实现

每个编码器层包含多头注意力和前馈网络：

class EncoderLayer(torch.nn.Module): def __init__(self, d_model, num_heads, d_ff, dropout=0.1): super().__init__() self.self_attn = MultiHeadAttention(d_model, num_heads) self.feed_forward = torch.nn.Sequential( torch.nn.Linear(d_model, d_ff), torch.nn.ReLU(), torch.nn.Linear(d_ff, d_model) ) self.norm1 = torch.nn.LayerNorm(d_model) self.norm2 = torch.nn.LayerNorm(d_model) self.dropout = torch.nn.Dropout(dropout) def forward(self, x, mask): attn_output = self.self_attn(x, x, x, mask) x = self.norm1(x + self.dropout(attn_output)) ff_output = self.feed_forward(x) return self.norm2(x + self.dropout(ff_output))

4.2 解码器层实现

解码器层比编码器复杂，包含两个注意力机制：

class DecoderLayer(torch.nn.Module): def __init__(self, d_model, num_heads, d_ff, dropout=0.1): super().__init__() self.self_attn = MultiHeadAttention(d_model, num_heads) self.src_attn = MultiHeadAttention(d_model, num_heads) self.feed_forward = torch.nn.Sequential( torch.nn.Linear(d_model, d_ff), torch.nn.ReLU(), torch.nn.Linear(d_ff, d_model) ) self.norm1 = torch.nn.LayerNorm(d_model) self.norm2 = torch.nn.LayerNorm(d_model) self.norm3 = torch.nn.LayerNorm(d_model) self.dropout = torch.nn.Dropout(dropout) def forward(self, x, memory, src_mask, tgt_mask): # 自注意力 attn_output = self.self_attn(x, x, x, tgt_mask) x = self.norm1(x + self.dropout(attn_output)) # 源注意力 attn_output = self.src_attn(x, memory, memory, src_mask) x = self.norm2(x + self.dropout(attn_output)) # 前馈网络 ff_output = self.feed_forward(x) return self.norm3(x + self.dropout(ff_output))

5. 训练与评估

5.1 训练循环实现

使用PyTorch 2.8的编译优化加速训练：

def train(model, iterator, optimizer, criterion, clip): model.train() epoch_loss = 0 for i, (src, trg) in enumerate(iterator): optimizer.zero_grad() output = model(src, trg[:, :-1]) output_dim = output.shape[-1] output = output.contiguous().view(-1, output_dim) trg = trg[:, 1:].contiguous().view(-1) loss = criterion(output, trg) loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), clip) optimizer.step() epoch_loss += loss.item() return epoch_loss / len(iterator)

5.2 评估结果分析

在测试集上评估模型性能：

def evaluate(model, iterator, criterion): model.eval() epoch_loss = 0 with torch.no_grad(): for i, (src, trg) in enumerate(iterator): output = model(src, trg[:, :-1]) output_dim = output.shape[-1] output = output.contiguous().view(-1, output_dim) trg = trg[:, 1:].contiguous().view(-1) loss = criterion(output, trg) epoch_loss += loss.item() return epoch_loss / len(iterator)