Installing PyTorch:
pip install torch
Setting up Transfomers:
Run the following command to clone the transformer repository, change to the desired version and install it.
git clone https://github.com/huggingface/transformers.git
cd transformers
git checkout 7b75aa9fa55bee577e2c7403301ed31103125a35
pip install -e .
Replace modeling_bart.py and modeling_utils.py inside src/transformers (of the library installed in the previous step) with the scripts in the bart directory.
python cli.py --do_train --output_dir out/${data}_checkpoint \
--checkpoint ${checkpoint_path} \
--train_file data/${data}/train.json \
--predict_file data/${data}/dev.json \
--train_batch_size ${train_bs} \
--predict_batch_size ${test_bs} \
--append_another_bos --do_lowercase
Use --checkpoint to specify the path to the checkpoint (if no path is passed, bart-large is loaded by default)
Use --train_file to specify the path to the training data.
Use -- predict_file to specify the path to the dev data.
Use --train_batch_size and --predict_batch_size to specify the training and prediction batch sizes.
The script will save the best checkpoint inside out/${data}_checkpoint.
Check cli.py for other command line arguments.
Run the following command for evaluation
python cli.py --do_predict --output_dir out/${data}_checkpoint \
--predict_file data/${data}/test.json \
--predict_batch_size ${test_bs} \
--append_another_bos --prefix test_
embeddings.json file uploaded at: https://drive.google.com/file/d/1KWiLaLPvEfRSPNQ0JDMA4Rx6uwtWaKsi/view?usp=sharing
Use --predict_file to specify the path to the file to be used for prediction. The above command uses the checkpoint out/${data}_checkpoint/best-model.pt . The predictions are saved inside out/${data}_checkpoint .