This is a repo for natural language inference using Match-LSTM and pre-trained BERT. This repo is based on https://github.com/donghyeonk/match-lstm.
SNLI 1.0 dataset will be automatically downloaded to data/ and unzipped if it doesn't exist.
Train the model with default parameters.
python main.py
Train the model on GPU 0 and 1.
python main.py --gpu=0,1
Train the model on GPU 0 and finetune Bert.
python main.py --gpu=0 --train_bert
Recommended training command:
python main.py --batch_size=512 --gpu=0
If you want to fine-tune Bert which increases the number of parameters significantly, you'd better reduce the batch size.
python main.py --batch_size=32 --gpu=0 --train_bert
or
python main.py --batch_size=64 --gpu=0,1 --train_bert