Awesome

CBA-online-CL

This is an official PyTorch implementation of "CBA: Improving Online Continual Learning via Continual Bias Adaptor" by Quanziang Wang, Renzhen Wang, Yichen Wu, Xixi Jia, and Deyu Meng.

Please contact Quanziang Wang (quanziangwang@gmail.com), Renzhen Wang (rzwang@xjtu.edu.cn), or Deyu Meng (dymeng@xjtu.edu.cn).

• We extended the Mammoth framework with our method and applied it to various rehearsal-based baselines.

Setup and Training

• Use python utils/main.py to run experiments.

• Results will be saved in ./results.

• Training examples:

  • Baseline DER++ on CIFAR-100 with buffer size M=2000

    python utils/main.py --model derpp --backbone resnet18 --dataset seq-cifar100 --lr 0.03 --batch_size 32 --minibatch_size 32 --n_epochs 1 --buffer_size 2000 --seed <5000> --gpu_id <0> --exp <onl-buf2000>

  • Our method DER-CBA on CIFAR-100 with buffer size M=2000

    python utils/main.py --model derpp_cba_online --backbone resnet18-meta --dataset seq-cifar100 --lr 0.03 --batch_size 32 --minibatch_size 32 --n_epochs 1 --buffer_size 2000 --seed <5000> --gpu_id <0> --exp <onl-buf2000>

  • We recommend repeating the experiment multiple times with different random seeds to reduce the effect of randomness, especially under the online setting (i.e., --n_epochs 1).

Requirements

• torch==1.7.0
• torchvision=0.9.0
• quadprog=0.1.7

Cite Our Work

If you find our work or this code is useful, please cite us:

@inproceedings{cba,
  title={CBA: Improving Online Continual Learning via Continual Bias Adaptor},
  author={Wang, Quanziang and Wang, Renzhen and Wu, Yichen and Jia, Xixi and Meng, Deyu},
  booktitle={ICCV},
  year={2023}
}