On Divergence Measures for Bayesian Pseudocoresets

This repository is the official implementation of On Divergence Measures for Bayesian Pseudocoresets (NeurIPS 2022)

Generating Expert Trajectories

Before training any Bayesian pseudocoreset, you'll need to generate some expert trajectories using buffer.py

python buffer.py --dataset=CIFAR10 --model=ConvNet --train_epochs=50 --num_experts=100

Training Bayesian pseudocoresets

The following command will then use the buffers we just generated to train Bayesian pseudocoresets of each divergence measure:

python train.py --dataset=CIFAR10 --model=ConvNet --divergence={fkl, rkl, wasserstein} --ipc={1, 10, 20} --eval_method={hmc, sghmc}

Acknowledgments

Our code is adapted from https://github.com/GeorgeCazenavette/mtt-distillation

Citation

If you find this useful in your research, please consider citing our paper:

@inproceedings{kim2022pseudocoresets,
  title     = {On Divergence Measures for Bayesian Pseudocoresets},
  author    = {Balhae Kim and Jungwon Choi and Seanie Lee and Yoonho Lee and Jung-Woo Ha and Juho Lee},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2022}
}