Hyperbolic Busemann Learning with Ideal Prototypes

This is the implementation of paper Hyperbolic Busemann Learning with Ideal Prototypes (NeurIPS2021).

How to use?

Ideal Prototype Learning

As the first step, you should learn ideal prototypes for the classes in the ultimate task. To run the code and learn the prototypes, the number of classes and output dimension should be specified.

To learn prototypes with 50 output dimensionality and 100 classes, use the code,

python prototype_learning.py -d 50 -c 100

The output will be prototypes-50d-100c.npy saved in prototypes directory.

Main code

Once the prototypes are ready, it's time to run the main code. To run HBL.py, the parameters in the argparser should be specified,

python HBL.py --data_name cifar100 -e 1110 -s 128 -r adam -l 0.0005 -c 0.00005 --mult 0.1 --datadir data/ --resdir runs/output_dir/cifar/ --hpnfile prototypes/prototypes-50d-100c.npy --logdir test --do_decay True --drop1 1000 --drop2 1100 --seed 100

Further explanation will be added soon.

Citation

Please consider citing this work using this BibTex entry,

@inproceedings{atigh2021hyperbolic,
  title={Hyperbolic Busemann Learning with Ideal Prototypes},
  author={Atigh, Mina Ghadimi and Keller-Ressel, Martin and Mettes, Pascal},
  booktitle={Thirty-Fifth Conference on Neural Information Processing Systems},
  year={2021}
}