LV-BERT

Introduction

In this repo, we introduce LV-BERT by exploiting layer variety for BERT. For detailed description and experimental results, please refer to our paper LV-BERT: Exploiting Layer Variety for BERT (Findings of ACL 2021).

Requirements

• Python 3.6
• TensorFlow 1.15
• numpy
• scikit-learn

Experiments

Firstly, set your data dir (absolute) to place datasets and models by

DATA_DIR=/path/to/data/dir

Fine-tining

We give the instruction to fine-tune a pre-trained LV-BERT-small (13M parameters) on GLUE. You can refer to this Google Colab notebook for a quick example. All models are provided in Google Drive, Huggingface and GitHub release. The models are pre-trained 1M steps with sequence length 128 to save compute. *_seq512 named models are trained for more 100K steps with sequence length 512 whichs are used for long-sequence tasks like SQuAD. See our paper for more details on model performance.

1. Create your data directory.

mkdir -p $DATA_DIR/models && cp vocab.txt $DATA_DIR/

Put the pre-trained model in the corresponding directory

mv lv-bert_small $DATA_DIR/models/

2. Download the GLUE data by running

python3 download_glue_data.py

3. Set up the data by running

cd glue_data && mv CoLA cola && mv MNLI mnli && mv MRPC mrpc && mv QNLI qnli && mv QQP qqp && mv RTE rte && mv SST-2 sst && mv STS-B sts && mv diagnostic/diagnostic.tsv mnli && mkdir -p $DATA_DIR/finetuning_data && mv * $DATA_DIR/finetuning_data && cd ..

4. Fine-tune the model by running

bash finetune.sh $DATA_DIR

PS: (a) You can test different tasks by changing configs in finetune.sh. (b) Some of the datasets on GLUE are small, causing that the results may vary substantially for different random seeds. The same as ELECTRA, we report the median of 10 fine-tuning runs from the same pre-trained model for each result.

Pre-training

We give the instruction to pre-train LV-BERT-small (13M parameters) using the OpenWebText corpus.

1. First download the OpenWebText pre-traing corpus (12G).

2. After downloading the pre-training corpus, build the pre-training dataset tf-record by running

bash build_data.sh $DATA_DIR

3. Then, pre-train the model by running

bash pretrain.sh $DATA_DIR

Bibtex

@inproceedings{yu2021lv-bert,
        author = {Yu, Weihao and Jiang, Zihang and Chen, Fei, Hou, Qibin and Feng, Jiashi},
        title = {LV-BERT: Exploiting Layer Variety for BERT},
        booktitle = {Findings of ACL},
        month = {August},
        year = {2021}
}

Reference

This repo is based on the repo ELECTRA.

Acknowledgment

Weihao Yu would like to thank TPU Research Cloud (TRC) program for the support of partial computational resources.