This repository contains an implementation with PyTorch of model presented in the paper "BERT-EMD: Many-to-Many Layer Mapping for BERT Compression with Earth Mover's Distance" in EMNLP 2020. The figure below illustrates a high-level view of the model's architecture. For more details about the techniques of BERT-EMD, refer to our paper.

Run command below to install the environment (using python3).

pip install -r requirements.txt 

1. Get GLUE data:

python download_glue_data.py --data_dir glue_data --tasks all

BaiduYun for alternative

2. Get BERT-Base offical model from here, download and unzip to directory ./model/bert_base_uncased. Convert tf model to pytorch model:

cd bert_finetune
python convert_bert_original_tf_checkpoint_to_pytorch.py \
--tf_checkpoint_path ../model/bert_base_uncased \
--bert_config_file ../model/bert_base_uncased/bert_config.json \
--pytorch_dump_path ../model/pytorch_bert_base_uncased

Or you can download the pytorch version directly from huggingface and download to ../model/pytorch_bert_base_uncased.

3. Get finetune teacher model, take task MRPC for example (working dir: ./bert_finetune):

export MODEL_PATH=../model/pytorch_bert_base_uncased/
export TASK_NAME=MRPC
python run_glue.py \
  --model_type bert \
  --model_name_or_path $MODEL_PATH \
  --task_name $TASK_NAME \
  --do_train \
  --do_eval \
  --do_lower_case \
  --data_dir ../data/glue_data/$TASK_NAME/ \
  --max_seq_length 128 \
  --per_gpu_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 4.0 \
  --save_steps 2000 \
  --output_dir ../model/$TASK_NAME/teacher/ \
  --evaluate_during_training \
  --overwrite_output_dir

4. Get the pretrained general distillation TinyBERT v2 student model: 4-layer and 6-layer. Unzip to directory model/student/layer4 and model/student/layer6 respectively. (This link may be temporarily unavailable, for alternative you can download from here BaiduYun).

5. Distill student model, take 4-layer student model for example:

cd ../bert_emd
export TASK_NAME=MRPC
python emd_task_distill.py  \
--data_dir ../data/glue_data/$TASK_NAME/ \
--teacher_model ../model/$TASK_NAME/teacher/ \
--student_model ../model/student/layer4/ \
--task_name $TASK_NAME \
--output_dir ../model/$TASK_NAME/student/ \
--beta 0.01 --theta 1

We replace the layer weight update method with division by addition. In our experiments, this normalization method is better than softmax on some datasets. Wight can be in range from 1e-3 to 1e 3

We add the hyperparameters for best-performing models as bellow and fixed some bugs.