Awesome

Decoupled Kullback-Leibler (DKL) Divergence Loss

This repository contains the implementation code for our arXiv paper Decoupled Kullback-Leibler (DKL) Divergence Loss, arXiv.

Overview

In this paper, we delve deeper into the KullbackLeibler (KL) Divergence loss and observe that it is equivalent to the Doupled Kullback-Leibler (DKL) Divergence loss that consists of 1) a weighted Mean Square Error (wMSE) loss and 2) a Cross4 Entropy loss incorporating soft labels. From our analysis of the DKL loss, we have identified two areas for improvement. Firstly, we address the limitation of DKL in scenarios like knowledge distillation by breaking its asymmetry property in training optimization. This modification ensures that the wMSE component is always effective during training, providing extra constructive cues. Secondly, we introduce global information into DKL for intra-class consistency regularization. With these two enhancements, we derive the Improved KullbackLeibler (IKL) Di11 vergence loss and evaluate its effectiveness by conducting experiments on CIFAR12 10/100 and ImageNet datasets, focusing on adversarial training and knowledge dis13 tillation tasks. The proposed approach achieves new state-of-the-art performance on both tasks, demonstrating the substantial practical merits.

Results and Pretrained Models for Knowledge Distillation

ImageNet

Method	Model-Teacher	Model-Student	Training Speed	Top-1 Acc(%)	link	log
ReviewKD	ResNet-34	ResNet18	0.319 s/iter	71.61	-	-
DKD	ResNet-34	ResNet18	-	71.70	-	-
IKL-KD	ResNet-34	ResNet18	0.197 s/iter	71.91	-	-

Method	Model-Teacher	Model-Student	Training Speed	Top-1 Acc(%)	link	log
ReviewKD	ResNet-50	MobileNet	0.526 s/iter	72.56	-	-
DKD	ResNet-50	MobileNet	-	72.05	-	-
IKL-KD	ResNet-50	MobileNet	0.252 s/iter	72.84	-	-

Results and Pretrained Models for Adversarial Robustness

By 2023/05/20, with IKL loss, we achieve new state-of-the-art adversarial robustness under settings that with/without augmentation strategies on auto-attack.

CIFAR-100 with autoaug

#	Method	Model	Natural Acc	Robust Acc (AutoAttack)	link	log
1	DAJAT	WRN-34-10	68.74	31.30	-	-
2	IKL-AT	WRN-34-10	65.93	32.52	model	log

CIFAR-100 with basic data preprocessing (random crop and random horizontal flip)

#	Method	Model	Natural Acc	Robust Acc (AutoAttack)	link	log
1	AWP	WRN-34-10	60.38	28.86	-	-
2	LBGAT	WRN-34-10	62.31	29.33	-	-
3	LAS-AT	WRN-34-10	62.99	30.77	-	-
4	ACAT	WRN-34-10	65.75	30.23	-	-
5	IKL-AT	WRN-34-10	66.51	31.43	model	log
6	IKL-AT	WRN-34-10	63.40	31.92	model	log

CIFAR-100 with synthesized data

#	Method	Model	Natural Acc	Robust Acc (AutoAttack)	link	log
1	Wang et al. (better diffusion models) 1M	WRN-28-10	68.06	35.65	-	-
2	Wang et al. (better diffusion models) 50M	WRN-28-10	72.58	38.83	-	-
3	IKL-AT 1M	WRN-28-10	68.99	35.89	-	-
4	IKL-AT 50M	WRN-28-10	73.85	39.18	model	log

CIFAR-10 with basic data preprocessing (random crop and random horizontal flip)

#	Method	Model	Natural Acc	Robust Acc (AutoAttack)	link	log
1	AWP	WRN-34-10	85.36	56.17	-	-
2	LBGAT	WRN-34-20	88.70	53.57	-	-
3	LAS-AT	WRN-34-10	87.74	55.52	-	-
4	ACAT	WRN-34-10	82.41	55.36	-	-
5	IKL-AT	WRN-34-10	85.31	57.13	model	log

CIFAR-10 with synthesized data

#	Method	Model	Natural Acc	Robust Acc (AutoAttack)	link	log
1	Wang et al. (better diffusion models) 1M	WRN-28-10	91.12	63.35	-	-
2	Wang et al. (better diffusion models) 20M	WRN-28-10	92.44	67.31	-	-
3	IKL-AT 1M	WRN-28-10	90.75	63.54	-	-
4	IKL-AT 20M	WRN-28-10	92.16	67.75	model	log

Training

More training scripts will be provided soon to reproduce our results on knowledge distillation and adversarial training tasks.

For the adversarial training task:
cd adv_training
bash sh/train_dkl_cifar100.sh
bash sh/train_dkl_cifar100_autoaug.sh
bash sh/train_dkl_cifar10.sh

Evaluation

before running the evaluation with auto-attack, please download the pre-trained models.

cd adv_training/auto_attacks
bash sh/eval.sh

Contact

If you have any questions, feel free to contact us through email (jiequancui@gmail.com) or Github issues. Enjoy!

BibTex

If you find this code or idea useful, please consider citing our related work:

@article{cui2023decoupled,
  title={Decoupled Kullback-Leibler Divergence Loss},
  author={Cui, Jiequan and Tian, Zhuotao and Zhong, Zhisheng and Qi, Xiaojuan and Yu, Bei and Zhang, Hanwang},
  journal={arXiv preprint arXiv:2305.13948},
  year={2023}
}

@inproceedings{cui2021learnable,
  title={Learnable boundary guided adversarial training},
  author={Cui, Jiequan and Liu, Shu and Wang, Liwei and Jia, Jiaya},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={15721--15730},
  year={2021}
}