BDR-main

This repository is an official PyTorch implementation of "Balanced Destruction-Reconstruction Dynamics for Memory-replay Class Incremental Learning"

Getting Started

This framework is implemented under Pytorch. It's better to implement an independent virtual environment on machine like this:

conda create --name pybdr python=3.9
conda activate pybdr
conda install pytorch=1.2.1
conda install torchvision -c pytorch

Download the Datasets

CIFAR-100

It will be downloaded automatically by torchvision when running the experiments.

ImageNet-Subset

You may download the dataset using the following links like this or directly from:

• Download from Google Drive

ImageNet

See the terms of ImageNet here.

Running Experiments

Before training, you need to set the dataset path _BASE_DATA_PATH in ./dataset/dataset_config.py.

Here is an example of how to use the code of BDR for training.

CUDA_VISIBLE_DEVICES=0 python main_incremental.py --exp-name nc_first_50_ntask_6 \
	--datasets cifar100_icarl --num-tasks 6 --nc-first-task 50 --network resnet18_cifar --seed 1993 \
	--nepochs 160 --batch-size 128 --lr 0.1 --momentum 0.9 --weight-decay 5e-4 --decay-mile-stone 80 120 \
	--clipping -1 --results-path results --save-models \
	--approach lucir_cwd_BDR --lamb 5.0 --num-exemplars-per-class 20 --exemplar-selection herding \
	--aux-coef 0.5 --reject-threshold 1 --dist 0.5 \
	--cwd --BDR --m1 0.8 --m2 0.8\

Citation

If you find this repository useful to your research, please consider citing:

@misc{zhou2023balanced,
      title={Balanced Destruction-Reconstruction Dynamics for Memory-replay Class Incremental Learning}, 
      author={Yuhang Zhou and Jiangchao Yao and Feng Hong and Ya Zhang and Yanfeng Wang},
      year={2023},
      eprint={2308.01698},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgement

This repository is built based on FACIL and CwD. We thank the authors for releasing their codes.