DAP-BERT

Installation

This repo is tested on Ubuntu 18.04 using PyTorch 1.6.0. Download the GLUE dataset using download_glue_data.py, i.e. run

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

Data augmentation

1. Get pre-trained bert model from huggingface.
2. Get GloVe word embeddings glove.42B.300d.txt.
3. Run

   python data_augmentation.py --pretrained_bert_model ${BERT_BASE_DIR} \
       --glove_embs ${GLOVE_EMB} \
       --glue_dir ${GLUE_DIR} \  
       --task_name ${TASK_NAME}
   

Prepare your fine-tuned model

Depending on the resource constraint, choose

1. fine-tuning a bert-base-uncased by scripts/fine_tune.sh, e.g.

   bash scripts/ft_bertbase.sh CoLA bert-base-uncased 10 42
   

2. or download fine-tuned models from TinyBERT, i.e. TinyBERT(4layer-312dim) or TinyBERT(6layer-768dim).

Please put all models in DAP-BERT/models.

Differentiable Architecture Pruning

Edit scripts/demo.sh to change paths to fine-tuned models and GLUE dataset. Example to run demo.sh:

bash scripts/demo.sh 0.5 0.1 CoLA 10 10 42

CoLA_search: searched model, with sub-architecture indices in arch.json

CoLA_ft: fine-tuning sub-network by intermediate distillation

CoLA_ft_pred: fine-tuning sub-network by prediction distillation