ConvBERT

Introduction

In this repo, we introduce a new architecture ConvBERT for pre-training based language model. The code is tested on a V100 GPU. For detailed description and experimental results, please refer to our NeurIPS 2020 paper ConvBERT: Improving BERT with Span-based Dynamic Convolution.

Requirements

• Python 3
• tensorflow 1.15
• numpy
• scikit-learn

Experiments

Pre-training

These instructions pre-train a medium-small sized ConvBERT model (17M parameters) using the OpenWebText corpus.

To build the tf-record and pre-train the model, download the OpenWebText corpus (12G) and setup your data directory in build_data.sh and pretrain.sh. Then run

bash build_data.sh

The processed data require roughly 30G of disk space. Then, to pre-train the model, run

bash pretrain.sh

See configure_pretraining.py for the details of the supported hyperparameters.

Fine-tining

We gives the instruction to fine-tune a pre-trained medium-small sized ConvBERT model (17M parameters) on GLUE. You can refer to the Google Colab notebook for a quick example. See our paper for more details on model performance. Pre-trained model can be found here. (You can also download it from baidu cloud with extraction code m9d2.)

To evaluate the performance on GLUE, you can download the GLUE data by running

python3 download_glue_data.py

Set up the data by running 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. After preparing the GLUE data, setup your data directory in finetune.sh and run

bash finetune.sh

And you can test different tasks by changing configs in finetune.sh.

If you find this repo helpful, please consider cite

@article{Jiang2020ConvBERT,
  title={ConvBERT: Improving BERT with Span-based Dynamic Convolution},
  author={Zi-Hang Jiang and Weihao Yu and Daquan Zhou and Y. Chen and Jiashi Feng and S. Yan},
  journal={ArXiv},
  year={2020},
  volume={abs/2008.02496}
}

References

Here are some great resources we benefit:

Codebase: Our codebase are based on ELECTRA.

Dynamic convolution: Implementation from Pay Less Attention with Lightweight and Dynamic Convolutions

Dataset: OpenWebText from Language Models are Unsupervised Multitask Learners