A GPU efficient vectorized Plackett-Luce loss in PyTorch from Language Modelling via Learning to Rank, Frydenlund et al. AAAI22.

1. This is tested for partial Plackett-Luce (k known ranks of v total classes), it is untested for full rankings.
2. This includes the biased partitioned preference modification, which treats all items in a partition (order) as equally preferred.

See the testing or training scripts for usage.

Please cite when using this code

@inproceedings{frydenlund2022language, title={Language Modelling via Learning to Rank}, author={Frydenlund, Arvid and Singh, Gagandeep and Rudzicz, Frank}, booktitle={Proceedings of the AAAI Conference on Artificial Intelligence}, volume={36}, number={10}, pages={10636--10644}, year={2022} }