Controllable Relation Disentanglement for Few-Shot Class-Incremental Learning

Our code and pretrained models are all provided in Google Drive, since Google Drive does not have much limitations on the size of files.

For testing the performance of our models, please download our code and run the following commands step-by-step.

🎯 Step-I: Download ctrl_fscil.zip from Google Drive and unzip it.

unzip ctrl_fscil.zip

🎯 Step-II: Prepare environments.

cd ctrl_fscil conda env create -f environment.yml && conda activate ctrl_fscil

🎯 Step-III: Prepare datasets.

cd data tar -xzvf ./cifar/cifar-100-python.tar.gz -C ./cifar tar -xvf miniimagenet.tar

🎯 Step-IV: Run evaluation scripts.

cd ../ bash run_ctrl/cifar/ctrl_inc.sh; bash run_ctrl/mini_imagenet/ctrl_inc.sh

Note that we run the code on an RTX 3090 GPU card with CUDA==11.4. We believe that CUDA==11.7 is also compatible with our code.

📌 If our work is useful for your research, please consider cite our paper:

  title={Controllable Relation Disentanglement for Few-Shot Class-Incremental Learning},
  author={Zhou, Yuan and Hong, Richang and Guo, Yanrong and Liu, Lin and Hao, Shijie and Zhang, Hanwang},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2025},
  publisher={IEEE}
}```