AdamB: Adam with Decoupled Bayes by Backprop

This is sample codes in AdamB: Decoupled Bayes by Backprop With Gaussian Scale Mixture Prior paper.

Requirements

This code has been tested on pytorch 1.9.0 with python 3.8.7

Example code execution

To train ResNet-18 by AdamB

$ cd examples
$ python train_adamb.py

ECE evaluation after train.

$ python inference_adamb.py

Note

Normalization layer runs Adam (not AdamB) without regularization.

Citation

If used in research, please cite AdamB by the following publications:

@ARTICLE{nishida2022adamb,
  author={Nishida, Keigo and Taiji, Makoto},
  journal={IEEE Access}, 
  title={AdamB: Decoupled Bayes by Backprop With Gaussian Scale Mixture Prior}, 
  year={2022},
  volume={10},
  number={},
  pages={92959-92970},
  doi={10.1109/ACCESS.2022.3203484}}

Acknowledgements