Patient Knowledge Distillation for BERT Model Compression

Knowledge distillation for BERT model


Run command below to install the environment

conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
pip install -r requirements.txt


Objective Function

L = (1 - \alpha) L_CE + \alpha * L_DS + \beta * L_PT,

where L_CE is the CrossEntropy loss, DS is the usual Distillation loss, and PT is the proposed loss. Please see our paper below for more details.

Data Preprocess

Modify the HOME_DATA_FOLDER in and put all data under it (by default it is ./data), RTE data is uploaded for your convenience.

  • The folder name under HOME_DATA_FOLDER should be
    • data_raw: store the raw datas of all tasks. So put downloaded raw data under here
      • MRPC
      • RTE
      • … (other tasks)
    • data_feat: store the tokenized data under this folder (optional)
      • MRPC
      • RTE
  • models
    • pretrained: put downloaded pretrained model (bert-base-uncased) under this folder

Predefinted Training

Run to start training, you can set DEBUG = True to run some pre-defined arguments

  • set argv = get_predefine_argv(‘glue’, ‘RTE’, ‘finetune_teacher’) or argv = get_predefine_argv(‘glue’, ‘RTE’, ‘finetune_student’) to start the normal fine-tuning
  • run to get teacher’s prediction for KD or PKD.
    • set output_all_layers = True for patient teacher
    • set output_all_layers = False for normal teacher
  • set argv = get_predefine_argv(‘glue’, ‘RTE’, ‘kd’) to start the vanilla KD
  • set argv = get_predefine_argv(‘glue’, ‘RTE’, ‘kd.cls’) to start the vanilla KD


This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.


If you find this code useful for your research, please consider citing:

title={Patient Knowledge Distillation for BERT Model Compression},
author={Sun, Siqi and Cheng, Yu and Gan, Zhe and Liu, Jingjing},
journal={arXiv preprint arXiv:1908.09355},

Paper is available at here.


View Github