NAPA-VQ: Neighborhood-Aware Prototype Augmentation with Vector Quantization for Continual Learning
We propose NAPA-VQ: Neighborhood Aware Prototype Augmentation with Vector Quantization, a framework that reduces catastrophic forgetting in Non-Exemplar based Class Incremental Learning.
Python 3.7.4
PyTorch: 1.9.0
Create the "dataset" directory under NAPA-VQ and download the following datasets into the created directory.
1.CIFAR-100
2.Tiny-ImageNet
3.ImageNet-Subset
CIFAR-100
# 5 Tasks
python main_cifar.py --fg_nc 50 --task_num 5 --custom_name "cifar_100-5" --base_model "cifar_100-5" --epochs 100 --shuffle > cifar-5.txt
# 10 Tasks
python main_cifar.py --fg_nc 50 --task_num 10 --custom_name "cifar_100-10" --base_model "cifar_100-10" --epochs 100 --shuffle > cifar-10.txt
# 20 Tasks
python main_cifar.py --fg_nc 40 --task_num 20 --custom_name "cifar_100-20" --base_model "cifar_100-20" --epochs 100 --shuffle > cifar-20.txt
Tiny-ImageNet
# 5 Tasks
python main_tiny.py --fg_nc 100 --task_num 5 --base_model "tiny-5" --custom_name "tiny-5" --shuffle --epochs 50 > tiny-5.txt
# 10 Tasks
python main_tiny.py --fg_nc 100 --task_num 10 --base_model "tiny-10" --custom_name "tiny-10" --shuffle --epochs 50 > tiny-10.txt
# 20 Tasks
python main_tiny.py --fg_nc 100 --task_num 20 --base_model "tiny-20" --custom_name "tiny-20" --shuffle --epochs 50 > tiny-20.txt
ImageNet-Subset
# 5 Tasks
python main_imagenet.py --custom_name "imagenet-5" --base_model "imagenet-5" --fg_nc 50 --task_num 5 --shuffle > imagenet-5.txt
# 10 Tasks
python main_imagenet.py --custom_name "imagenet-10" --base_model "imagenet-10" --fg_nc 50 --task_num 10 --shuffle > imagenet-10.txt
# 20 Tasks
python main_imagenet.py --custom_name "imagenet-20" --base_model "imagenet-20" --fg_nc 40 --task_num 20 --shuffle > imagenet-20.txt
@InProceedings{Malepathirana_2023_ICCV,
author = {Malepathirana, Tamasha and Senanayake, Damith and Halgamuge, Saman},
title = {NAPA-VQ: Neighborhood-Aware Prototype Augmentation with Vector Quantization for Continual Learning},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2023},
pages = {11674-11684}
}
We thank the authors of the following repositories for their excellent codebase providing reusable functions.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent