Continual Learning With Filter Atom Swapping

Pytorch Implementation of Continual Learning With Filter Atom Swapping (ICLR'22 Spolight) Paper

If you find this repo helpful for your work, please cite the following bib entries:

@inproceedings{Miao_2022_CLatoms,
title={Continual Learning With Filter Atom Swapping},
author={Miao, Zichen and Wang, Ze and Chen, Wei and Qiu, Qiang},
booktitle={International Conference on Learning Representations},
year={2022}}

Requirements

• PyTorch >= version 1.8

Class-incremental experiments

Dataset

CIFAR100 will be automatically downloaded to 'Datasets' dir. Please download ImageNet to 'Datasets/ImageNet'

Training Scripts

cifar100 (N=5)

$ ./train_dcf_c100.sh 6 GPU_TO_USE

cifar100 (N=10)

$ ./train_dcf_c100.sh 11 GPU_TO_USE

ImageNet-Sub (N=5)

$ ./train_dcf_img100.sh 6 GPU_TO_USE

ImageNet-Sub (N=10)

$ ./train_dcf_img100.sh 11 GPU_TO_USE

Results