TutorKD

This repository is about Tutor-KD long paper: Tutoring Helps Students Learn Better: Improving Knowledge Distillation for BERT with Tutor Network published in EMNLP 2022. In this project, we are interested in generating the traning samples which can mitigate the incorrect teacher predictions and repetitive learning for student.

Overview

Training

Main Results

Requirements

• Python 3
• Transformers 4.2.2
• Numpy
• pytorch

Quick start

Prepare the pre-training corpora(Wikipedia and Bookcorpus) in data folder. Use python preprocess.py.

• --data_path: A directory containing pre-processed examples (pickle file).
• --raw_data_path: A directory containing raw text examples.

Distillation

Finally, use python distillation.py for distillation.

• --config: A Student model architecture. Choose model architecture from: half, extreme-12, ext-6, ext-2
• --lr: Set the learning rate.
• --epochs: Set the number of epochs.
• --batch_size: Set the batch size for conducting at once.
• --step_batch_size: Set the batch size for updating per each step (If the memory of GPU is enough, set the batch_size and step_batch_size the same.
• --data_path: A directory containing pre-processed examples.
• --model_save_path: Set the directory for saving the student model

Contact Info

For help or issues using Tutor-KD, please submit a GitHub issue.

For personal communication related to Tutor-KD, please contact Junho Kim <[email protected]>.