BERT预训练实战从零构建WikiText-2双任务模型1. 环境准备与数据加载在开始BERT预训练之前我们需要搭建合适的开发环境并准备数据集。本实验使用PyTorch框架和WikiText-2数据集这是一个包含维基百科文章的经典语言建模数据集大小适中适合教学演示。首先安装必要的Python包pip install torch transformers datasets matplotlib加载WikiText-2数据集并进行预处理from datasets import load_dataset dataset load_dataset(wikitext, wikitext-2-raw-v1) print(f数据集结构: {dataset}) # 过滤空行和标题行 train_texts [text for text in dataset[train][text] if len(text) 0 and not text.startswith( )]2. 构建BERT模型架构我们将实现一个简化版的BERT模型包含以下核心组件import torch import torch.nn as nn from transformers import BertConfig class BertForPretraining(nn.Module): def __init__(self, config): super().__init__() self.bert BertModel(config) self.mlm_head nn.Linear(config.hidden_size, config.vocab_size) self.nsp_head nn.Linear(config.hidden_size, 2) def forward(self, input_ids, attention_mask, token_type_ids): outputs self.bert(input_ids, attention_mask, token_type_ids) sequence_output outputs.last_hidden_state pooled_output outputs.pooler_output mlm_logits self.mlm_head(sequence_output) nsp_logits self.nsp_head(pooled_output) return mlm_logits, nsp_logits配置模型参数简化版config BertConfig( vocab_size30522, hidden_size128, # 原始BERT为768 num_hidden_layers2, # 原始BERT为12 num_attention_heads2, intermediate_size256, max_position_embeddings512 ) model BertForPretraining(config)3. 数据预处理与特征工程BERT预训练需要特殊的数据处理流程包括WordPiece分词和双任务标签生成。3.1 构建词汇表与分词器from tokenizers import Tokenizer, models, trainers tokenizer Tokenizer(models.WordPiece(unk_token[UNK])) trainer trainers.WordPieceTrainer( special_tokens[[PAD], [UNK], [CLS], [SEP], [MASK]], vocab_sizeconfig.vocab_size ) tokenizer.train_from_iterator(train_texts, trainer)3.2 生成训练样本import random def create_mlm_sample(text): tokens tokenizer.encode(text).ids # 随机mask 15%的token masked_indices random.sample(range(len(tokens)), kint(0.15*len(tokens))) labels [-100] * len(tokens) for idx in masked_indices: # 80%概率替换为[MASK] if random.random() 0.8: masked_token tokenizer.token_to_id([MASK]) # 10%概率替换为随机token elif random.random() 0.5: masked_token random.randint(0, config.vocab_size-1) # 10%概率保持不变 else: masked_token tokens[idx] labels[idx] tokens[idx] tokens[idx] masked_token return tokens, labels def create_nsp_sample(text_a, text_b): # 50%概率选择连续句子或随机句子 if random.random() 0.5: text_b random.choice(train_texts) label 0 # NotNext else: label 1 # IsNext return text_a, text_b, label4. 训练过程实现4.1 损失函数定义BERT预训练包含两个任务的联合损失criterion_mlm nn.CrossEntropyLoss(ignore_index-100) criterion_nsp nn.CrossEntropyLoss() def compute_loss(mlm_logits, nsp_logits, mlm_labels, nsp_labels): mlm_loss criterion_mlm( mlm_logits.view(-1, config.vocab_size), mlm_labels.view(-1) ) nsp_loss criterion_nsp(nsp_logits, nsp_labels) return mlm_loss nsp_loss4.2 训练循环from torch.utils.data import DataLoader from tqdm import tqdm optimizer torch.optim.AdamW(model.parameters(), lr5e-5) for epoch in range(5): model.train() total_loss 0 for batch in tqdm(train_loader): optimizer.zero_grad() input_ids batch[input_ids].to(device) attention_mask batch[attention_mask].to(device) token_type_ids batch[token_type_ids].to(device) mlm_labels batch[mlm_labels].to(device) nsp_labels batch[nsp_labels].to(device) mlm_logits, nsp_logits model(input_ids, attention_mask, token_type_ids) loss compute_loss(mlm_logits, nsp_logits, mlm_labels, nsp_labels) loss.backward() optimizer.step() total_loss loss.item() print(fEpoch {epoch1} | Avg Loss: {total_loss/len(train_loader):.4f})5. 损失曲线分析与调优训练过程中记录两种任务的损失值import matplotlib.pyplot as plt plt.figure(figsize(10, 5)) plt.plot(mlm_losses, labelMLM Loss) plt.plot(nsp_losses, labelNSP Loss) plt.xlabel(Steps) plt.ylabel(Loss) plt.title(BERT Pretraining Loss Curves) plt.legend() plt.grid() plt.show()典型损失曲线特征分析阶段MLM Loss特征NSP Loss特征可能原因初期快速下降波动较大模型学习基础表征中期平稳下降逐渐收敛模型捕获语言规律后期小幅波动基本稳定接近局部最优解关键调优参数建议# 推荐参数配置 optimal_config { batch_size: 32, # 平衡显存和稳定性 max_seq_length: 128, # 短文本效率更高 learning_rate: 3e-5, # BERT经典初始学习率 warmup_steps: 1000, # 学习率预热 weight_decay: 0.01 # 防止过拟合 }6. 模型验证与应用6.1 掩码语言预测示例def predict_masked_token(text, mask_position): inputs tokenizer(text, return_tensorspt) inputs.input_ids[0, mask_position] tokenizer.mask_token_id with torch.no_grad(): logits model(**inputs).logits predicted_token_id logits[0, mask_position].argmax().item() return tokenizer.decode(predicted_token_id) text The capital of France is [MASK]. print(predict_masked_token(text, mask_position5)) # 输出可能: paris6.2 下一句关系判断def predict_next_sentence(sentence_a, sentence_b): inputs tokenizer(sentence_a, sentence_b, return_tensorspt) with torch.no_grad(): outputs model(**inputs) prob torch.softmax(outputs.logits, dim-1) return prob[0, 1].item() # IsNext的概率 sentence_a The weather is nice today. sentence_b Lets go to the park. print(fIsNext概率: {predict_next_sentence(sentence_a, sentence_b):.2f}) # 输出可能: 0.877. 进阶优化技巧7.1 动态掩码策略原始BERT在预处理时静态mask tokens现代实现采用动态maskingdef dynamic_masking(tokens): # 每个epoch重新生成mask模式 masking_strategy random.choice([ token_level, span_level, whole_word ]) # 实现不同的mask模式... return masked_tokens7.2 梯度累积在显存有限时使用梯度累积模拟更大batch sizeaccumulation_steps 4 for i, batch in enumerate(train_loader): loss compute_loss(batch) loss loss / accumulation_steps loss.backward() if (i1) % accumulation_steps 0: optimizer.step() optimizer.zero_grad()7.3 混合精度训练使用FP16加速训练from torch.cuda.amp import GradScaler, autocast scaler GradScaler() for batch in train_loader: optimizer.zero_grad() with autocast(): loss compute_loss(batch) scaler.scale(loss).backward() scaler.step(optimizer) scaler.update()8. 实际应用建议资源有限时减小hidden_size至256或层数至4仍能保持不错效果领域适应在专业领域语料上继续预训练Domain-Adaptive Pretraining损失震荡当NSP损失波动较大时可尝试降低学习率或增大batch size早停策略当验证集loss连续3个epoch不下降时停止训练完整训练50步后的典型损失值参考任务初始Loss最终Loss下降幅度MLM8.52.175%NSP0.690.3549%通过本实战教程我们实现了BERT核心预训练流程观察到双任务损失的有效收敛。虽然简化版模型参数较少但完整保留了BERT的核心训练机制为后续微调任务奠定了良好基础。