• 以文本分类为例

六、Trainer和文本分类

---

一、导入相关包

!pip install transformers datasets evaluate accelerate

from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments
from datasets import load_dataset

---

二、加载数据集

dataset = load_dataset("csv", data_files="./ChnSentiCorp_htl_all.csv", split="train")
dataset = dataset.filter(lambda x: x["review"] is not None)
dataset
'''
Dataset({
    features: ['label', 'review'],
    num_rows: 7765
})
'''

---

三、划分数据集

datasets = dataset.train_test_split(test_size=0.1)
datasets
'''
DatasetDict({
    train: Dataset({
        features: ['label', 'review'],
        num_rows: 6988
    })
    test: Dataset({
        features: ['label', 'review'],
        num_rows: 777
    })
})
'''

---

四、数据集预处理

import torch

tokenizer = AutoTokenizer.from_pretrained("hfl/rbt3")

def process_function(examples):
    tokenized_examples = tokenizer(examples["review"], max_length=128, truncation=True)
    tokenized_examples["labels"] = examples["label"]
    return tokenized_examples

tokenized_datasets = datasets.map(process_function, batched=True, 
                                  remove_columns=datasets["train"].column_names)
tokenized_datasets
'''
DatasetDict({
    train: Dataset({
        features: ['input_ids', 'token_type_ids', 'attention_mask', 'labels'],
        num_rows: 6988
    })
    test: Dataset({
        features: ['input_ids', 'token_type_ids', 'attention_mask', 'labels'],
        num_rows: 777
    })
})
'''

---

五、创建模型（区别一）

def model_init():
    model = AutoModelForSequenceClassification.from_pretrained("hfl/rbt3")
    return model

---

六、创建评估函数

import evaluate

acc_metric = evaluate.load("accuracy")
f1_metirc = evaluate.load("f1")

def eval_metric(eval_predict):
    predictions, labels = eval_predict
    predictions = predictions.argmax(axis=-1)
    acc = acc_metric.compute(predictions=predictions, references=labels)
    f1 = f1_metirc.compute(predictions=predictions, references=labels)
    acc.update(f1)
    return acc

---

七、创建 TrainingArguments(区别二)

• logging_steps=500为了防止多次训练 log 太多可以增大 logging_steps

train_args = TrainingArguments(output_dir="./checkpoints",      # 输出文件夹
                               per_device_train_batch_size=64,  # 训练时的batch_size
                               per_device_eval_batch_size=128,  # 验证时的batch_size
                               logging_steps=500,               # log 打印的频率
                               evaluation_strategy="epoch",     # 评估策略
                               save_strategy="epoch",           # 保存策略
                               save_total_limit=3,              # 最大保存数
                               learning_rate=2e-5,              # 学习率
                               weight_decay=0.01,               # weight_decay
                               metric_for_best_model="f1",      # 设定评估指标
                               load_best_model_at_end=True)     # 训练完成后加载最优模型

---

八、创建 Trainer(区别三)

• 没有指定 model而是指定 model_init

from transformers import DataCollatorWithPadding
trainer = Trainer(model_init=model_init, 
                  args=train_args, 
                  train_dataset=tokenized_datasets["train"], 
                  eval_dataset=tokenized_datasets["test"], 
                  data_collator=DataCollatorWithPadding(tokenizer=tokenizer),
                  compute_metrics=eval_metric)


# 之前
from transformers import DataCollatorWithPadding
trainer = Trainer(model=model,
                  args=train_args,
                  train_dataset=tokenized_datasets["train"],
                  eval_dataset=tokenized_datasets["test"],
                  data_collator=DataCollatorWithPadding(tokenizer=tokenizer),
                  compute_metrics=eval_metric)

---

九、模型训练

trainer.train()

---

十、模型训练(自动搜索)(区别四)

!pip install optuna

• 使用默认的超参数空间
• compute_objective=lambda x: x["eval_f1"]中的 x是指的评价函数的返回值,在这里因为没有显示的指定评价函数返回值的 key,所以 f1的 key采用默认值 eval_f1

trainer.hyperparameter_search(compute_objective=lambda x: x["eval_f1"], direction="maximize", n_trials=10)

• 自定义超参数空间
  • 可以在default_hp_space_optuna 函数中增加 trainer 的选项

def default_hp_space_optuna(trial):
    return {
        "learning_rate": trial.suggest_float("learning_rate", 1e-6, 1e-4, log=True),
        "num_train_epochs": trial.suggest_int("num_train_epochs", 1, 5),
        "seed": trial.suggest_int("seed", 1, 40),
        "per_device_train_batch_size": trial.suggest_categorical("per_device_train_batch_size", [4, 8, 16, 32, 64]),
        "optim": trial.suggest_categorical("optim", ["sgd", "adamw_hf"]),
    }

trainer.hyperparameter_search(hp_space=default_hp_space_optuna, compute_objective=lambda x: x["eval_f1"], direction="maximize", n_trials=10)

---

启动 tensorboard

• 进入运行日志文件夹
  • 终端启动

!tensorboard --logdir runs

• jupyter 启动

# 运行这行代码将加载 TensorBoard并允许我们将其用于可视化
%reload_ext tensorboard 
%tensorboard --logdir=./runs/

---