Repository files navigation

PyTorch implementation of BERT in "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding" (https://arxiv.org/abs/1810.04805)

• Python 3.6+
• PyTorch 4.1+
• tqdm

All dependencies can be installed via:

pip install -r requirements.txt

First things first, you need to prepare your data in an appropriate format.Your corpus is assumed to follow the below constraints.

• Each line is adocument.
• Adocument consists ofsentences, seperated by vertical bar (|).
• Asentence is assumed to be already tokenized. Tokens are seperated by space.
• Asentence has no more than 256 tokens.
• Adocument has at least 2 sentences.
• You have two distinct data files, one for train data and the other for val data.

This repo comes with example data for pretraining in data/example directory.Here is the content of data/example/train.txt file.

One, two, three, four, five,|Once I caught a fish alive,|Six, seven, eight, nine, ten,|Then I let go again.I’m a little teapot|Short and stout|Here is my handle|Here is my spout.Jack and Jill went up the hill|To fetch a pail of water.|Jack fell down and broke his crown,|And Jill came tumbling after.

Also, this repo includes SST-2 data in data/SST-2 directory for sentiment classification.

python bert.py preprocess-index data/example/train.txt --dictionary=dictionary.txt

Running the above command produces dictionary.txt file in your current directory.

python bert.py pretrain --train_data data/example/train.txt --val_data data/example/val.txt --checkpoint_output model.pth

This step trains BERT model with unsupervised objective. Also this step does:

• logs the training procedure for every epoch
• outputs model checkpoint periodically
• reports the best checkpoint based on validation metric

You can fine-tune pretrained BERT model with downstream task.For example, you can fine-tune your model with SST-2 sentiment classification task.

python bert.py finetune --pretrained_checkpoint model.pth --train_data data/SST-2/train.tsv --val_data data/SST-2/dev.tsv

This command also logs the procedure, outputs checkpoint, and reports the best checkpoint.

• Transformer-pytorch : My own implementation of Transformer. This BERT implementation is based on this repo.

@dreamgonfly