Sinkhorn Distance Minimization for Knowledge Distillation (COLING 2024)

Installation

To install the environment, run:

sh ins.sh

Download GLUE and SuperGLUE Data

Download the GLUE data using this repository or from GLUE benchmark website, unpack it to directory datas/glue and rename the folder CoLA to COLA.

Download the SuperGLUE data from SuperGLUE benchmark website.

Download Pre-trained BERT

Download bert_uncased_L-12_H-768_A-12 (BERT-base) and bert_uncased_L-6_H-768_A-12 for teacher model and student model, respectively, from this repository. and use the API from Huggingface to transform them to pytorch checkpoint.

Task-specific BERT Model Distillation

The training script for Task-specific Teacher Model Finetuning can be found in the script/teacher/ directory, where $TEACHER_PATH denotes the file path of the teacher model.

Similarly, the training script for Task-specific Student Model Distillation is located in the script/student/ directory. In this case, $STUDENT_PATH and $TEACHER_PATH represent the file paths of the student and teacher models, respectively.

Task-specific T0 Model Distillation

To install the environment, run:

sh T0/ins.sh

To perform Task-specific Teacher Model Finetuning, run:

python3 T0/distillation_t.py --dataset_name super_glue --dataset_config_name DATASET_NAME --template_name "TEMPLATE_NAME" --model_name_or_path MODEL_DIR --output_dir ./debug --parallelize

To perform Task-specific Student Model Distillation, run:

python3 T0/distillation.py --dataset_name super_glue --dataset_config_name DATASET_NAME --template_name "TEMPLATE_NAME" --model_name_or_path MODEL_DIR --output_dir ./debug --parallelize

Task-specific GPT Model Distillation

To install the environment, run:

sh GPT-Neo/ins.sh

To perform Task-specific Teacher Model Finetuning, run:

python3 GPT-Neo/distillation_t.py --dataset_name super_glue --dataset_config_name DATASET_NAME --template_name "TEMPLATE_NAME" --model_name_or_path MODEL_DIR --output_dir ./debug --parallelize

To perform Task-specific Student Model Distillation, run:

python3 GPT-Neo/distillation.py --dataset_name super_glue --dataset_config_name DATASET_NAME --template_name "TEMPLATE_NAME" --model_name_or_path MODEL_DIR --output_dir ./debug --parallelize

Student Checkpoints

The distilled student model for each task reported in the paper can be downloaded using the following link: https://drive.google.com/drive/folders/1BsA0VHKSa_-Bp5I7dQ2Ftk2q7cIyPrdC

Teacher Checkpoints

The teacher model for each task reported in the paper can be downloaded using the following link: https://drive.google.com/file/d/1sBi35Dk8VJ7TU0warB6BL9QKx-in9Ww6/view?usp=drive_link

BibTeX

@article{cui2024sinkhorn,
  title={Sinkhorn Distance Minimization for Knowledge Distillation},
  author={Cui, Xiao and Qin, Yulei and Gao, Yuting and Zhang, Enwei and Xu, Zihan and Wu, Tong and Li, Ke and Sun, Xing and Zhou, Wengang and Li, Houqiang},
  journal={arXiv preprint arXiv:2402.17110},
  year={2024}
}