Official implementation of the paper [New Insights on Relieving Task-Recency Bias for Online Class Incremental Learning] (TCSVT 2023).
The backbone of project mainly refers to online-continual-learning.
Create a virtual enviroment
virtualenv online-cl
Activating a virtual environment
source online-cl/bin/activate
Installing packages
pip install -r requirements.txt
Except our implementation code, you could easily find other implementation results from SCR, DVC, ER-ACE and OCM.
Detailed descriptions of options can be found in general_main.py.
You can run python file "run_cifar10.py", "run_cifar100.py" and "run_mini.py" to reimplement our paper results, for example:
python run_mini.py
Detailed commands are as follows:
python general_main.py --agent er --loss rfocal --classify max --data cifar10 --eps_mem_batch 100 --mem_size 200 --review_trick True --kd_trick True --kd_lamda 0.05 --cor_prob 0.99 --T 20.0 --fix_order True
python general_main.py --agent er --loss rfocal --classify max --data cifar10 --eps_mem_batch 100 --mem_size 500 --review_trick True --kd_trick True --kd_lamda 0.05 --cor_prob 0.99 --T 20.0 --fix_order True
python general_main.py --agent er --loss rfocal --classify max --data cifar10 --eps_mem_batch 100 --mem_size 1000 --review_trick True --kd_trick True --kd_lamda 0.1 --cor_prob 0.99 --T 20.0 --fix_order True
python general_main.py --agent er --loss rfocal --classify max --data cifar100 --eps_mem_batch 100 --mem_size 1000 --review_trick True --kd_trick True --kd_lamda 0.15 --cor_prob 0.99 --T 20.0 --fix_order True
python general_main.py --agent er --loss rfocal --classify max --data cifar100 --eps_mem_batch 100 --mem_size 2000 --review_trick True --kd_trick True --kd_lamda 0.05 --cor_prob 0.99 --T 20.0 --fix_order True
python general_main.py --agent er --loss rfocal --classify max --data cifar100 --eps_mem_batch 100 --mem_size 5000 --review_trick True --kd_trick True --kd_lamda 0.1 --cor_prob 0.99 --T 20.0 --fix_order True
python general_main.py --agent er --loss rfocal --classify max --data mini_imagenet --eps_mem_batch 100 --mem_size 1000 --review_trick True --kd_trick True --kd_lamda 0.05 --cor_prob 0.99 --T 20.0 --fix_order True
python general_main.py --agent er --loss rfocal --classify max --data mini_imagenet --eps_mem_batch 100 --mem_size 2000 --review_trick True --kd_trick True --kd_lamda 0.1 --cor_prob 0.99 --T 20.0 --fix_order True
python general_main.py --agent er --loss rfocal --classify max --data mini_imagenet --eps_mem_batch 100 --mem_size 5000 --review_trick True --kd_trick True --kd_lamda 0.05 --cor_prob 0.99 --T 20.0 --fix_order True
If you use this paper/code in your research, please consider citing us:
New Insights on Relieving Task-Recency Bias for Online Class Incremental Learning
@ARTICLE{10287323,
author={Liang, Guoqiang and Chen, Zhaojie and Chen, Zhaoqiang and Ji, Shiyu and Zhang, Yanning},
journal={IEEE Transactions on Circuits and Systems for Video Technology},
title={New Insights on Relieving Task-Recency Bias for Online Class Incremental Learning},
year={2023},
volume={},
number={},
pages={1-1},
doi={10.1109/TCSVT.2023.3325651}}
Other traditional papers we encourage you to cite can be found in RaptorMai.
Thanks RaptorMai for selflessly sharing his implementation about recent state-of-the-art methods.