Stochastic dual coordinate ascent for training conditional random fields. Visit the project webpage for more details.
Python 3.6, Numpy, Scipy, Matplotlib, tensorboard_logger.
Call main.py with the desired arguments.
The full list of arguments is specified in sdca4crf/arguments.py.
The main training loop is in sdca4crf/sdca.py.
A typical use case is:
python main.py --dataset ner --non-uniformity 0.8 --sampling-scheme gap
You can use tensorboard to visualize training. Training curves and other results are also saved into pickle files at the end of training.
Four pre-processed datasets are available under data/.
To use another dataset, you should extract features and numberize them.
The folders named experiments contain a bunch of scripts used for the paper.