---

前言

在目前的各项NLP任务中，如果要在特定任务或者领域应用文本分类，数据分布一定是有一些差距的。这时候可以考虑进行深度预训练。进行继续预训练，有利于提升任务的性能。

一、继续预训练是什么？

根据基于bert模型的下游NLP任务的特定语料，对模型进行领域的继续训练。主要类型如下：

Within-task pre-training：Bert在训练语料上进行预训练

In-domain pre-training：在同一领域上的语料进行预训练

Cross-domain pre-training：在不同领域上的语料进行预训练
 

二、代码

1.引入库

代码如下（示例）：

import numpy as np
import pandas as pd
import warnings
warnings.filterwarnings('ignore')
from transformers import (AutoModelForMaskedLM,AutoTokenizer, LineByLineTextDataset,                           DataCollatorForLanguageModeling,                           Trainer, TrainingArguments)

2.读入数据

代码如下（示例）：

train_data = pd.read_csv('data/train/train.csv', sep='\t') 
test_data = pd.read_csv('data/test/test.csv', sep='\t') 
train_data['text'] = train_data['title'] + '.' + train_data['abstract'] 
test_data['text'] = test_data['title'] + '.' + test_data['abstract'] 
data = pd.concat([train_data, test_data]) 
data['text'] = data['text'].apply(lambda x: x.replace('\n', ''))  
text = '\n'.join(data.text.tolist())

3.构建数据集

代码如下（示例）：

model_name = 'roberta-base'
model = AutoModelForMaskedLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.save_pretrained('./paper_roberta_base')

train_dataset = LineByLineTextDataset(
    tokenizer=tokenizer,
    file_path="text.txt",  # mention train text file here
    block_size=256)

valid_dataset = LineByLineTextDataset(
    tokenizer=tokenizer,
    file_path="text.txt",  # mention valid text file here
    block_size=256)

data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer, mlm=True, mlm_probability=0.15)

---

4.开始训练

training_args = TrainingArguments(
    output_dir="./paper_roberta_base_chk",  # select model path for checkpoint
    overwrite_output_dir=True,
    num_train_epochs=5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    gradient_accumulation_steps=2,
    evaluation_strategy='steps',
    save_total_limit=2,
    eval_steps=200,
    metric_for_best_model='eval_loss',
    greater_is_better=False,
    load_best_model_at_end=True,
    prediction_loss_only=True,
    report_to="none")

trainer = Trainer(
    model=model,
    args=training_args,
    data_collator=data_collator,
    train_dataset=train_dataset,
    eval_dataset=valid_dataset)

trainer.train()
trainer.save_model(f'./paper_roberta_base')

总结

面对一些NLP比赛，可以考虑进行深度预训练。