Source code and dataset for "TR-BERT: Dynamic Token Reduction for Accelerating BERT Inference".
The code is based on huggaface's transformers. Thanks to them!
Install dependencies and apex:
pip3 install -r requirement.txt
pip3 install --editable transformers
Download the DistilBERT-3layer and BERT-1024 from Google Drive/Tsinghua Cloud.
Download the IMDB, Yelp, 20News datasets from Google Drive/Tsinghua Cloud.
Download the Hyperpartisan dataset, and randomly split it into train/dev/test set: python3 split_hyperpartisan.py
Use flag --do train:
python3 run_classification.py --task_name imdb --model_type bert --model_name_or_path bert-base-uncased --data_dir imdb --max_seq_length 512 --per_gpu_train_batch_size 8 --per_gpu_eval_batch_size 16 --gradient_accumulation_steps 4 --learning_rate 3e-5 --save_steps 2000 --num_train_epochs 5 --output_dir imdb_models/bert_base --do_lower_case --do_eval --evaluate_during_training --do_train
where task_name can be set as imdb/yelp_f/20news/hyperpartisan for different tasks and model type can be set as bert/distilbert for different models.
Use flag --do_eval_grad.
python3 run_classification.py --task_name imdb --model_type bert --model_name_or_path imdb_models/bert_base --data_dir imdb --max_seq_length 512 --per_gpu_train_batch_size 8 --per_gpu_eval_batch_size 8 --output_dir imdb_models/bert_base --do_lower_case --do_eval_grad
This step doesn't supoort data DataParallel or DistributedDataParallel currently and should be done in a single GPU.
Start from the checkpoint from the task-specific fine-tuned model. Change model_type from bert to autobert, and run with flag --do_train --train_rl:
python3 run_classification.py --task_name imdb --model_type autobert --model_name_or_path imdb_models/bert_base --data_dir imdb --max_seq_length 512 --per_gpu_train_batch_size 8 --per_gpu_eval_batch_size 8 --gradient_accumulation_steps 4 --learning_rate 3e-5 --save_steps 2000 --num_train_epochs 3 --output_dir imdb_models/auto_1 --do_lower_case --do_train --train_rl --alpha 1 --guide_rate 0.5
where alpha is the harmonic coefficient for the length punishment and guide_rate is the proportion of imitation learning steps. model_type can be set as autobert/distilautobert for applying token reduction to BERT/DistilBERT.
Use flag --do_eval_logits.
python3 run_classification.py --task_name imdb --model_type bert --model_name_or_path imdb_models/bert_base --data_dir imdb --max_seq_length 512 --per_gpu_train_batch_size 8 --per_gpu_eval_batch_size 8 --output_dir imdb_models/bert_base --do_lower_case --do_eval_logits
This step doesn't supoort data DataParallel or DistributedDataParallel currently and should be done in a single GPU.
Start from the checkpoint from --train_rl model and run with flag --do_train --train_both --train_teacher:
python3 run_classification.py --task_name imdb --model_type autobert --model_name_or_path imdb_models/auto_1 --data_dir imdb --max_seq_length 512 --per_gpu_train_batch_size 8 --per_gpu_eval_batch_size 1 --gradient_accumulation_steps 4 --learning_rate 3e-5 --save_steps 2000 --num_train_epochs 3 --output_dir imdb_models/auto_1_both --do_lower_case --do_train --train_both --train_teacher --alpha 1
Use flag --do_eval:
python3 run_classification.py --task_name imdb --model_type autobert --model_name_or_path imdb_models/auto_1_both --data_dir imdb --max_seq_length 512 --per_gpu_train_batch_size 8 --per_gpu_eval_batch_size 1 --output_dir imdb_models/auto_1_both --do_lower_case --do_eval --eval_all_checkpoints
When the batch size is more than 1 in evaluating, we will remain the same number of tokens for each instance in the same batch.
For IMDB dataset, we find that when we directly initialize the selector with heuristic objective before train
the policy network solely, we can get a bit better performance. For other datasets, this step makes little change. Run this step with flag --do_train --train_init:
python3 trans_imdb_rank.py
python3 run_classification.py --task_name imdb --model_type initbert --model_name_or_path imdb_models/bert_base --data_dir imdb --max_seq_length 512 --per_gpu_train_batch_size 8 --per_gpu_eval_batch_size 8 --gradient_accumulation_steps 4 --learning_rate 3e-5 --save_steps 2000 --num_train_epochs 3 --output_dir imdb_models/bert_init --do_lower_case --do_train --train_init
Download the SQuAD 2.0 dataset.
Download the MRQA dataset with our split] from Google Drive/Tsinghua Cloud.
Download the HotpotQA dataset from the Transformer-XH repository, where paragraphs are retrieved for each question according to TF-IDF, entity linking and hyperlink and re-ranked by BERT re-ranker.
Download the TriviaQA dataset, where paragraphs are re-rank by the linear passage re-ranker in DocQA.
Download the WikiHop dataset.
The whole training progress of question answer models is similiar to text classfication models, with flags --do_train, --do_train --train_rl, --do_train --train_both --train_teacher in turn. The codes of each dataset:
SQuAD: run_squad.py with flag version_2_with_negative
NewsQA / NaturalQA: run_mrqa.py
RACE: run_race_classify.py
HotpotQA: run_hotpotqa.py
TriviaQA: run_triviaqa.py
WikiHop: run_wikihop.py
The example harmonic coefficients are shown as follows:
| Dataset | train_rl | train_both |
|---|---|---|
| SQuAD 2.0 | 5 | 5 |
| NewsQA | 3 | 5 |
| NaturalQA | 2 | 2 |
| RACE | 0.5 | 0.1 |
| YELP.F | 2 | 0.5 |
| 20News | 1 | 1 |
| IMDB | 1 | 1 |
| HotpotQA | 0.1 | 4 |
| TriviaQA | 0.5 | 1 |
| Hyperparisan | 0.01 | 0.01 |
If you use the code, please cite this paper:
@inproceedings{ye2021trbert,
author = {Deming Ye and
Yankai Lin and
Yufei Huang and
Maosong Sun},
editor = {Kristina Toutanova and
Anna Rumshisky and
Luke Zettlemoyer and
Dilek Hakkani{-}T{\"{u}}r and
Iz Beltagy and
Steven Bethard and
Ryan Cotterell and
Tanmoy Chakraborty and
Yichao Zhou},
title = {{TR-BERT:} Dynamic Token Reduction for Accelerating {BERT} Inference},
booktitle = {Proceedings of the 2021 Conference of the North American Chapter of
the Association for Computational Linguistics: Human Language Technologies,
{NAACL-HLT} 2021, Online, June 6-11, 2021},
pages = {5798--5809},
publisher = {Association for Computational Linguistics},
year = {2021},
url = {https://doi.org/10.18653/v1/2021.naacl-main.463},
doi = {10.18653/v1/2021.naacl-main.463},
timestamp = {Fri, 06 Aug 2021 00:41:32 +0200},
biburl = {https://dblp.org/rec/conf/naacl/YeLHS21.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}