Feiertag is an open-source neural sequence tagging toolkit built with PyTorch.
Use the package manager pip to install feiertag.
git clone [email protected]:erip/feiertag.git
cd feiertag
pip install -e .Feiertag supports a handful of baseline architectures and datasets with configuration handled by Hydra. An example to train an NER model on CoNLL 2003 formatted data is shown below:
feiertag-train data.train="$DATA_DIR/train.txt" data.valid="$DATA_DIR/valid.txt" data_format=conll2003 trainer.max_epochs=25 embedding.path="$EMBEDDING_DIR/glove.6B.50d.txt" model=bilstm_crf data.loader.batch_size=128 embedding.freeze=falseModel training is handled by PyTorch-Lightning and kwargs passed to the pytorch_lightning.Trainer class can be provided by Hydra override syntax. For example, to emulate pl.Trainer(gpus='0,1', deterministic=True), use the following syntax:
feiertag-train ... +trainer.gpus="0,1" +trainer.deterministic=trueBy default, Hydra maintains an output directory structure to separate runs. Similarly, PyTorch-Lightning logs checkpoints, hyperparameters, and Tensorboard logfiles with the following structure:
outputs
└── 2020-12-02
└── 18-02-11
├── lightning_logs
│ └── version_0
│ ├── checkpoints
│ │ └── epoch=1.ckpt
│ ├── events.out.tfevents.1606950132.erip.16601.0
│ └── hparams.yaml
└── train.log
To view the above experiment in tensorboard, issue tensorboard --logdir outputs/2020-12-02/18-02-11/lightning_logs/version_0/.
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
Please make sure to update tests as appropriate.