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 a document.
- A document consists of sentences, seperated by vertical bar (|).
- A sentence is assumed to be already tokenized. Tokens are seperated by space.
- A sentence has no more than 256 tokens.
- A document 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.