/fine-tuning-Bert-for-sentiment-analysis

基于bert-base-chinese微调的中文情感分析任务，在WeiboSenti100k 数据集上训练5个epoch并且收敛

Primary LanguagePython

基于bert微调的中文情感分析

简介

  • 使用Bert-base-Chinese，对微博评论数据集进行情感分类，分别是[积极/消极]。

Change Logs

  • [2023/7/27] The script train.py supports parse arguments. Now you can run the train.py in CLI with the arguments you set

Requirements

pip install -r requirements.txt

Dataset

使用WeiboSenti100k数据集，该数据集包含在./dataset。 该数据集包含10万条**微博帖子，每条帖子都被标记为正面或负面。

Usage

  • train

    run the CLI

        python train.py <Args>

    or

        accelerate launch train.py <Args>

    Args

    • -o, --output_checkpoints=OUTPUT_CHECKPOINTS Default: './checkpoints' "the dir where you want to save checkpoints ."
    • -m, --model_path=MODEL_PATH Default: 'bert-base-chinese' "name of huggingface repo or a local model dir"
    • -d, --dataset_path=DATASET_PATH Default: './dataset/weibo_se... "the dataset dir"
    • -l, --learning_rate=LEARNING_RATE Default: 2e-05
    • -b, --batch_size=BATCH_SIZE Default: 80
    • -e, --epoch=EPOCH Default: 5
    • --weight_decay=WEIGHT_DECAY Default: 0.02
    • --warmup_ratio=WARMUP_RATIO Default: 0.2
    • -u, --use_gpu=USE_GPU Default: '0' "the indexes of gpus you want to use. such as "0,1,2","0","1,3,6" etc."

  • inference

        python inference.py -s "There input your sentence." # Inference for single sentence.
        python inference.py -i True # Continuous inference.

Related Docs