In the top level directory multitasking_bert/, run:
pip install .For NER tasks, prepare dataset using NERDataset.create_ner_dataset()
in multi_tasking_transformers/data/ner_dataset.py.
For the training scripts to work also for evaluation, partition the dataset into train, test, dev. The training script will create dataloaders and the multitasker will evaluation on the dev set during training and the test set at an interval specified in configuration (see Configuration).
config.gin shows an example of configuration of multitasking runs (see also config_STL.gin for single task runs).
Parameters include the path to preprocessed data, pre-trained model weights, and other model parameters.
multitask_train.py and singletask_train.py show examples of a training run using a biomedical dataset collection
containing separate tasks for entities and datasets.
To run training scripts:
python multitask_train.py