Introduction

This repo includes my customized and frequently used code framework used for a wide range of NLP experiments and some awesome(I think) functions in tutils.py. Updating continuously...

The framework includes main function, dataset, dataloader, model definition, train/valide loop and so on.

The main structure refers to the code of DAG-ERC, which is quite clear.

main dependency libs

This framework mainly includes some libs below:

• Pytorch 1.14
• Python 3.9
• transformers 4.23.1
• scikit-learn

Customized functionalities

• Linear learning rate scheduler
• separate learning rate training(separate params group)
• flexible metric selection provided by sklearn
• detailed log provided by logging library

Example

The framework uses finetune bert-base-uncased on SST-2(downloaded by hugging face datasets) as a demo

The best output is as below:

[2022-12-16 18:15:02,690][run.py][line:194][INFO] Epoch: 2, train_loss: 0.1376, train_acc: 95.1, train_fscore: 95.1, valid_loss: 0.2641, valid_acc: 92.43, valid_fscore: 92.43, test_loss: nan, test_acc: nan, test_fscore: 0.0, time: 131.87 sec