tks0123456789 / kaggle-Gendered_Pronoun_Resolution

Kaggle Gendered Pronoun Resolution

10th place solution for Gendered Pronoun Resolution on Kaggle

I borrow some code from pytorch-pretrained-BERT/example and [PyTorch] BERT Baseline (Public Score ~ 0.54).

Instruction

Build docker image using docker/Dockerfile.

Training

./train.sh offsets 30
./train.sh nli 30

It will save 90 models(111GB). If your GPU have only 8GB of RAM, change from

--train_batch_size 4 --gradient_accumulation_steps 5

to

--train_batch_size 1 --gradient_accumulation_steps 20

in train.sh.

Prediction

python offsets_predict.py --test_file "test_stage_2.tsv"
python nli_predict.py --test_file "test_stage_2.tsv"
python make_submission.py