使用simpletransformers库进行新闻文本多分类

183****0229

1、环境信息

pytorch 1.8.1
python 3.8.1
cuda 11.1

2、数据集信息

TNEWS’ 今日头条中文新闻（短文本）分类 Short Text Classificaiton for News

该数据集来自今日头条的新闻版块，共提取了15个类别的新闻，包括旅游，教育，金融，军事等。

     数据量：训练集(53,360)，验证集(10,000)，测试集(10,000)
     例子：
     {"label": "102", "label_des": "news_entertainment", "sentence": "江疏影甜甜圈自拍，迷之角度竟这么好看，美吸引一切事物"}
     每一条数据有三个属性，从前往后分别是 分类ID，分类名称，新闻字符串（仅含标题）。

TNEWS’数据集下载

3、下载代码并安装依赖环境

# 切换路径
cd /hy-tmp
# 代码已经上传到github
git clone https://github.com/JunnYu/hy_tutorial.git 
# 如果下载失败请使用镜像地址git clone https://hub.fastgit.org/JunnYu/hy_tutorial.git
# 切换路径
cd hy_tutorial
# 解压缩文件
unzip tnews_classfication.zip
# 切换路径
cd tnews_classfication
# 安装所需的python依赖包
pip install -r requirements.txt

4、运行程序

python train.py

import pandas as pd
from sklearn.metrics import accuracy_score
from simpletransformers.classification import ClassificationModel
# 读取tsv文件
train_df = pd.read_csv("dataset/processed/train.tsv", sep="\t")
eval_df = pd.read_csv("dataset/processed/dev.tsv", sep="\t")
# 定义label
label_list = ['news_story',
              'news_culture',
              'news_entertainment',
              'news_sports',
              'news_finance',
              'news_house',
              'news_car',
              'news_edu',
              'news_tech',
              'news_military',
              'news_travel',
              'news_world',
              'news_stock',
              'news_agriculture',
              'news_game']

model = ClassificationModel(
    model_type="bert",
    model_name="hfl/chinese-roberta-wwm-ext",
    num_labels=15,
    args={
        "reprocess_input_data": True,
        "overwrite_output_dir": True,
        "use_cached_eval_features": True,
        "do_lower_case": True,
        "evaluate_during_training": True,
        "labels_list": label_list,
        "max_seq_length": 256,
        "manual_seed": 42,
        "num_train_epochs": 10,
        "train_batch_size": 64,
        "eval_batch_size": 128,
        "save_optimizer_and_scheduler": False
    }
)

# 训练模型
model.train_model(train_df=train_df, eval_df=eval_df,
                  accuracy=accuracy_score)

# 评估模型
result, model_outputs, wrong_predictions = model.eval_model(
    eval_df,  accuracy=accuracy_score)
print(result)

可以打开train.py修改num_train_epochs为自己想要的论数。

5、查看GPU使用率

watch -n 0.1 nvidia-smi

6、查看开发集效果

进入outputs文件夹，双击training_progress_scores.csv或eval_results.txt

最好结果accuracy 66.85

7、Reference

https://github.com/ThilinaRajapakse/simpletransformers
https://github.com/CLUEbenchmark/CLUE