README.md

Topic classifier for biomedical articles

Multilabel topic classifier for medical articles.

This system learns a topic classifier based for articles labelelled with multiple topics. The included model uses a variant of BERT pre-trained on medical texts, and finetunes it on task instances.

Data

Input data is expected to be a json-formatted file containing a list of articles. Each article should have a title, an abstract and a topics field containing a list of topics.

Installing

virtualenv -p python3 env
source env/bin/activate
pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html

Training

python trainier.py [options]

optional arguments:
  -h, --help            show this help message and exit
  --gpus GPUS
  --nodes NODES
  --name NAME
  --fast_dev_run
  --train_filename TRAIN_FILENAME
  --learning_rate LEARNING_RATE
  --batch_size BATCH_SIZE
  --epochs EPOCHS
  --valid_size VALID_SIZE
  --max_len MAX_LEN
  --bert_flavor BERT_FLAVOR
  --selected_features SELECTED_FEATURES

Example training command line:

python trainer.py --gpus=-1 --name test1 --train_filename ../scrappers/data/20200529/litcovid.json

pytorch-lightning provides a tensorboard logger. You can check it with

tensorboard --logdir lightning_logs

Then point your browser to http://localhost:6006/.

Generating predictions

predict.py --checkpoint checkpoints/epoch\=0-val_loss\=0.2044.ckpt --test_filename ../scrappers/data/20200529/cord19-metadata.json > predicted.json