Mamba-FSCIL: Dynamic Adaptation with Selective State Space Model for Few-Shot Class-Incremental Learning

This is the official repository for "Mamba-FSCIL: Dynamic Adaptation with Selective State Space Model for Few-Shot Class-Incremental Learning," accepted by IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2026.

Mamba-FSCIL: Dynamic Adaptation with Selective State Space Model for Few-Shot Class-Incremental Learning
Xiaojie Li^1,2, Yibo Yang^4, Jianlong Wu^1,3, Yue Yu^2, Ming-Hsuan Yang^5, Liqiang Nie^1,3, Min Zhang^1,3
^1Harbin Institute of Technology (Shenzhen), ^2Pengcheng Laboratory, ^3Shenzhen Loop Area Institute, ^4King Abdullah University of Science and Technology (KAUST), ^5University of California, Merced

🔨 Installation

Follow these steps to set up your environment:

• Create and activate a new Conda environment:

  conda create --name mambafscil python=3.10 -y
  conda activate mambafscil

• Install CUDA and cuDNN: Follow the official CUDA installation instructions.

• Install PyTorch and torchvision:

  • Using pip:

    pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113

  • Using conda:

    conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch

• Install MMCV, OpenCV, and other dependencies:

  pip install -U openmim
  mim install mmcv-full==1.7.0
  pip install opencv-python matplotlib einops rope timm==0.6.12 scikit-learn==1.1.3 yapf==0.40.1
  git clone https://github.com/state-spaces/mamba.git; cd mamba; git checkout v1.2.0.post1; pip install .

• Clone the repository and set up the directory:

  git clone https://github.com/xiaojieli0903/Mamba-FSCIL.git
  cd Mamba-FSCIL; mkdir ./data

➡️ Data Preparation

• Download datasets from this link provided by NC-FSCIL.

• Organize the datasets in the ./data folder:

  --data
    ----cifar/
    ----CUB_200_2011/
    ----miniimagenet/

🚀 Training

Execute the provided scripts to start training.

CIFAR

sh train_cifar.sh

Session	0	1	2	3	4	5	6	7	8
Mamba-FSCIL	82.95±0.13	77.87±0.16	74.12±0.37	69.67±0.30	66.75±0.17	63.62±0.13	61.44±0.13	59.80±0.33	57.52±0.23

[Base Log] [Incremental Log]

Mini Imagenet

sh train_miniimagenet.sh

Session	0	1	2	3	4	5	6	7	8
Mamba-FSCIL	84.54±0.15	79.23±0.14	74.68±0.90	71.42±0.60	68.85±0.31	65.89±0.29	62.77±0.10	61.01±0.32	59.29±0.15

[Base Log] [Incremental Log]

CUB

sh train_cub.sh

Session	0	1	2	3	4	5	6	7	8	9	10
Mamba-FSCIL	80.92±0.04	76.28±0.07	73.03±0.08	70.12±0.03	67.77±0.12	65.72±0.06	65.29±0.13	64.05±0.07	62.35±0.04	62.14±0.06	61.52±0.11

[Base Log] [Incremental Log]

✏️ Citation

If you find our work useful in your research, please consider citing:

@article{li2026mamba,
  title={Mamba-FSCIL: Dynamic Adaptation with Selective State Space Model for Few-Shot Class-Incremental Learning},
  author={Li, Xiaojie and Yang, Yibo and Wu, Jianlong and Yu, Yue and Yang, Ming-Hsuan and Nie, Liqiang and Zhang, Min},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2026}
}

👍 Acknowledgments

This codebase builds on FSCIL. Thank you to all the contributors.