07agarg/haf

Code for the Paper Learning Hierarchy Aware Features for Reducing Mistake Severity, accepted in ECCV 2022

HAF in PyTorch

Official Code Release for Learning Hierarchy Aware Features for Reducing Mistake Severity
Ashima Garg, Depanshu Sani, Saket Anand.

European Conference on Computer Vision (ECCV 2022)

Citations

If you find this paper useful, please cite our paper:

@inproceedings{garg2022learning,
  title={Learning Hierarchy Aware Features for Reducing Mistake Severity},
  author={Garg, Ashima and Sani, Depanshu and Anand, Saket},
  booktitle={European Conference on Computer Vision},
  pages={252--267},
  year={2022},
  organization={Springer}
}

Proposed Approach

Installation

Clone this repository

$ git clone https://github.com/07Agarg/HAF.git
$ cd HAF

Dataset Preparation

Refer to Repository: Making Better Mistakes

Hierarchies

Refer to Repository: Making Better Mistakes

Using the Code for Training

The experiments in the paper are contained in the folder experiments/ dataset wise. For CIFAR-100

bash experiments/train/cifar-100/cross-entropy.sh

For iNaturalist-19

bash experiments/train/inat/cross-entropy.sh

For tiered-imagenet

bash experiments/train/tieredimagenet/cross-entropy.sh

For testing

Refer to the code repository: CRM-making-better-mistakes

Link To Trained Models

Download HAF final trained models from the link and use the above repository for testing.

Acknowledgements

This codebase is borrowed from making-better-mistakes

Contact

If you have any suggestion or question, you can leave a message here or contact us directly at [email protected]

haf's Issues

why do you use F.normalize() twice?

HAF/util/arch/custom_resnet18.py

Line 497 in fb29e36

classifier_1_weight = F.normalize(self.classifier_1[0].weight, p=2, dim=1)

HAF/util/arch/custom_resnet18.py

Line 542 in fb29e36

classifier_1_weight = F.normalize(classifier_1_weight, p=2, dim=1)

•    classifier_1_weight = F.normalize(self.classifier_1[0].weight, p=2, dim=1)
  

•    classifier_1_weight = F.normalize(classifier_1_weight, p=2, dim=1) 
  

I think it is not necessary there, twice