jewelc92/self-kd-lib

[ECCV-2022] Official implementation of MixSKD: Self-Knowledge Distillation from Mixup for Image Recognition && Pytorch Implementations of some Self-Knowledge Distillation and data augmentation methods

This project provides the implementations of some data augmentation methods, regularization methods, online Knowledge distillation and Self-Knowledge distillation methods.

Installation

Requirements

Ubuntu 18.04 LTS

Python 3.8 (Anaconda is recommended)

CUDA 11.1

PyTorch 1.12 + torchvision 0.13

Perform experiments on CIFAR-100 dataset

Dataset

CIFAR-100 : download

unzip to the ./data folder

The commands for running various methods can be found in main_cifar.sh

Top-1 accuracy(%) of Self-KD and Data Augmentation (DA) methods on ResNet-18
Type	Method	Venue	Accuracy(%)
Baseline	Cross-entropy	-	76.24
Self-KD	DDGSD [1]	AAAI-2019	76.61
	DKS [2]	CVPR-2019	78.64
	SAD [3]	ICCV-2019	76.40
	BYOT [4]	ICCV-2019	77.88
	Tf-KD-reg [5]	CVPR-2020	76.61
	CS-KD [6]	CVPR-2020	78.66
	FRSKD [7]	CVPR-2021	76.60
	PS-KD [8]	ICCV-2021	79.31
	BAKE [9]	arXiv:2104.13298	76.93
	MixSKD [10]	ECCV-2022	80.32
DA	Label Smoothing [1]	CVPR-2016	78.72
	Virtual Softmax [2]	NeurIPS-2018	78.54
	Focal Loss [3]	ICCV-2017	76.19
	Maximum Entropy [4]	ICLR Workshops 2017	76.50
	Cutout [5]	arXiv:1708.04552	76.66
	Random Erase [6]	AAAI-2020	76.75
	Mixup [7]	ICLR-2018	78.68
	CutMix [8]	ICCV-2019	80.17
	AutoAugment [9]	CVPR-2019	77.97
	RandAugment [10]	CVPR Workshops-2020	76.86
	AugMix [11]	arXiv:1912.02781	76.22
	TrivalAugment [12]	ICCV-2021	76.03

Some implementations are referred by the official code. Thanks the papers' authors for their released code. The results are reproduced by our released code, so it may be not strictly consistent with the original papers.

Perform experiments on ImageNet dataset

MixSKD	Top-1 Accuracy	Script	Log	Pretrained Model
ResNet-50	78.76	sh	Baidu Cloud	Baidu Cloud

Perform experiments of downstream object detection on COCO

Our implementation of object detection is based on MMDetection. Please refer the detailed guideline at https://github.com/winycg/detection.

Framework	mAP	Log	Pretrained Model
Cascade-Res50	41.6	Baidu Cloud	Baidu Cloud

Perform experiments on downstream semantic segmentation

The training script are based on our previous released segmentation codebase: https://github.com/winycg/CIRKD

Dataset	mIoU	Script	Log	Pretrained Model
ADE20K	42.37	sh	Baidu Cloud	Baidu Cloud
COCO-Stuff-164K	37.12	sh	Baidu Cloud	Baidu Cloud
Pascal VOC	78.78	sh	Baidu Cloud	Baidu Cloud

If you find this repository useful, please consider citing the following paper:

@inproceedings{yang2022mixskd,
  title={MixSKD: Self-Knowledge Distillation from Mixup for Image Recognition},
  author={Yang, Chuanguang and An, Zhulin and Zhou, Helong and  Cai, Linhang and Zhi, Xiang and Wu, Jiwen and Xu, Yongjun and Zhang, Qian},
  booktitle={European Conference on Computer Vision},
  year={2022}
}