raburabu91 / distill-bert

Knowledge Distillation from BERT

Distilled BERT

This work aims Knowledge Distillation from Google BERT model to compact Convolutional Models. (Not done yet)

Requirements

Python > 3.6, fire, tqdm, tensorboardx, tensorflow (for loading checkpoint file)

Example Usage

Fine-tuning (MRPC) Classifier with Pre-trained Transformer

Download BERT-Base, Uncased and GLUE Benchmark Datasets before fine-tuning.

• make sure that "total_steps" in train.json should be greater than n_epochs*(num_data/batch_size)

Modify several config json files before following commands for training and evaluating.

python finetune.py config/finetune/mrpc/train.json
python finetune.py config/finetune/mrpc/eval.json

Training Blend CNN from scratch

See Transformer to CNN. Modify several config json files before following commands for training and evaluating.

python classify.py config/blendcnn/mrpc/train.json
python classify.py config/blendcnn/mrpc/eval.json

Knowledge Distillation from finetuned Transformer to CNN

Modify several config json files before following commands for training and evaluating.

python distill.py config/distill/mrpc/train.json
python distill.py config/distill/mrpc/eval.json