ERD for IML

code for our CVPRW 2022 paper Incremental Meta-Learning via Episodic Replay Distillation for Few-Shot Image Recognition by Kai Wang, Xialei Liu, Andrew D. Bagdanov, Luis Herranz, Shangling Jui, and Joost van de Weijer.

Our supplementary material is also attached here as supp.pdf.

required packages

All installed packages in our running environment are in requirements.txt, please check whether you have any conflicts if enconterring any problem.

1st step: download dataset and create split

1,Pleas run

get_cifar_data.py

to download cifar dataset from torchvision.datasets automatically and it will be saved at ./data/cifar.

2, Run

cd ./data/cifar
./data/cifar/create_cifar_split_auto.py

to create the 16-task cifar split.

2nd step: run the experiments

cd back to the root directory. Then run

cd ../..
reproduce.sh

to reproduce the CIFAR100 experiments in our paper.

FT_cifar.py is the finetuning baseline. ERD_cifar.py is our method ERD implementation.

3rd step: test the experiments

Run

test_cifar.sh

to test the model performance.

By default, we print out the test accuracy and save the results in ./results/ directory.

REMIND ME if you need further information and details on other datasets since I may forget to update the repositories later.

Our bibtex is:

@InProceedings{Wang_2022_CVPR,
    author    = {Wang, Kai and Liu, Xialei and Bagdanov, Andrew D. and Herranz, Luis and Jui, Shangling and van de Weijer, Joost},
    title     = {Incremental Meta-Learning via Episodic Replay Distillation for Few-Shot Image Recognition},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2022},
    pages     = {3729-3739}
}

wangkai930418/ERD