ScrollNet: Dynamic Weight Importance for Continual Learning

Introduction

The official Pytorch implementation for ScrollNet: Dynamic Weight Importance for Continual Learning, Visual Continual Learning workshop, ICCV 2023.

Installation

Clone this github repository

git clone https://github.com/FireFYF/ScrollNet.git
cd ScrollNet

Create a conda enviroment

conda env create --file env.yml --name ScrollNet

Notice: set the appropriate version of your CUDA driver for cudatoolkit in env.yml.

Environment activation/deactivation

conda activate ScrollNet
conda deactivate

Launch experiments

Run with ScrollNet-FT

python -u src/main_incremental.py --gpu 0 --approach finetuning --results-path ./results/5splits/scrollnet_ft --num-tasks 5

Run with ScrollNet-LWF

python -u src/main_incremental.py --gpu 0 --approach lwf --results-path ./results/5splits/scrollnet_lwf --num-tasks 5

Run with ScrollNet-EWC

python -u src/main_incremental.py --gpu 0 --approach ewc --results-path ./results/5splits/scrollnet_ewc --num-tasks 5

Tune the number of subnetworks

Please modify the file 'SizeOfSubnetworks.yml'. The default setting is for 4 subnetworks with equal splitting (ScrollNet-4).

Acknowledgement

The implementation is based on FACIL, which was developed as a framework based on class-incremental learning. We suggest referring to it if you want to incorporate more CL methods into ScrollNet.

Cite

If you find this work useful for your research, please cite:

@misc{yang2023scrollnet,
    title={ScrollNet: Dynamic Weight Importance for Continual Learning},
    author={Fei Yang and Kai Wang and Joost van de Weijer},
    year={2023},
    eprint={2308.16567},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}