from datasets import load_dataset
# 下载 YelpReviewFull 数据集dataset = load_dataset("yelp_review_full")

from transformers import AutoTokenizer
# 加载预训练的 BERT Tokenizertokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
def tokenize_function(examples):"""使用 Tokenizer 对文本进行编码，并进行填充和截断"""return tokenizer(examples["text"], padding="max_length", truncation=True)
# 对数据集进行预处理tokenized_datasets = dataset.map(tokenize_function, batched=True)

# 从数据集中抽样 1000 个训练样本small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000))
# 从数据集中抽样 1000 个测试样本small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000))

from transformers import AutoModelForSequenceClassification
# 加载 BERT 模型，并设置标签数量为 5（情感评分从 1 到 5）model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)

from transformers import TrainingArguments
model_dir = "models/bert-base-cased-finetune-yelp"
# 配置训练参数training_args = TrainingArguments(output_dir=model_dir,# 模型保存路径per_device_train_batch_size=16,# 每个设备的训练批次大小num_train_epochs=5,# 训练轮数logging_steps=100# 每 100 步记录一次日志)

import numpy as npimport evaluate
# 加载准确率指标metric = evaluate.load("accuracy")
def compute_metrics(eval_pred):"""计算准确率"""logits, labels = eval_predpredictions = np.argmax(logits, axis=-1)# 将 logits 转换为预测值    return metric.compute(predictions=predictions, references=labels)

from transformers import Trainer
# 实例化 Trainertrainer = Trainer(model=model,args=training_args,train_dataset=small_train_dataset,# 训练数据集eval_dataset=small_eval_dataset,# 验证数据集compute_metrics=compute_metrics # 计算指标的函数)

# 更新训练参数配置training_args = TrainingArguments(output_dir=model_dir,evaluation_strategy="epoch",# 每个 epoch 结束时进行评估per_device_train_batch_size=16,num_train_epochs=3,logging_steps=30# 每 30 步记录一次日志)

# 开始训练trainer.train()

watch -n 1 nvidia-smi

# 保存训练后的模型trainer.save_model(model_dir)
# 保存训练状态trainer.save_state()

# 导入必要的库from datasets import load_datasetfrom transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainerimport numpy as npimport evaluate
# 数据集下载dataset = load_dataset("yelp_review_full")
# 数据预处理tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
def tokenize_function(examples):"""使用 Tokenizer 对文本进行编码，并进行填充和截断"""return tokenizer(examples["text"], padding="max_length", truncation=True)
tokenized_datasets = dataset.map(tokenize_function, batched=True)
# 数据抽样small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000))small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000))
# 模型加载与训练配置model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)
model_dir = "models/bert-base-cased-finetune-yelp"
training_args = TrainingArguments(output_dir=model_dir,per_device_train_batch_size=16,num_train_epochs=5,logging_steps=100)
# 指标评估metric = evaluate.load("accuracy")
def compute_metrics(eval_pred):"""计算准确率"""logits, labels = eval_predpredictions = np.argmax(logits, axis=-1)return metric.compute(predictions=predictions, references=labels)
# 实例化 Trainertrainer = Trainer(model=model,args=training_args,train_dataset=small_train_dataset,eval_dataset=small_eval_dataset,compute_metrics=compute_metrics)
# 开始训练trainer.train()
# 监控 GPU 使用# 使用命令行工具: watch -n 1 nvidia-smi
# 保存模型和训练状态trainer.save_model(model_dir)trainer.save_state()