MICS

MICS: Midpoint Interpolation to Learn Compact and Separated Representations for Few-Shot Class-Incremental Learning

MICS: Midpoint Interpolation to Learn Compact and Separated Representations for Few-Shot Class-Incremental Learning [Paper] [Supp]

Solang Kim, Yuho Jeong, Joon Sung Park, Sung Whan Yoon

In WACV 2024.

Installation

1. Clone this repository.

   git clone http://github.com/solang/mics.git
   

2. Install the required dependency.

   conda create env -y -n mics python=3.9
   conda activate mics
   bash install.sh
   

   • Recommend: Check your CUDA version using the nvcc -V command and update the torch version in the install.sh script accordingly. You can find the compatible PyTorch versions for your CUDA release at this link

   • Our codes are tested on Ubuntu 18.04 with Python 3.9.5 and Pytorch 1.9.0. We utilized NVIDIA RTX A5000 for mini-ImageNet (CUDA 10.1) and GeForce RTX 3090 for CIFAR-100 and CUB-200-2011 (CUDA 11.1)

Datasets

Download FSCIL benchmark datasets.

1. CIFAR100: https://www.cs.toronto.edu/~kriz/cifar.html
2. mini-ImageNet: There is no official website for mini-ImageNet. You can utilize the learn2learn python package or the unofficial Google Drive links for the download.
3. CUB-200-2011: https://www.vision.caltech.edu/datasets/cub_200_2011/

Train

cd scripts
bash [dataset]-mics.sh 

• Before execution the scripts, set your dataset path and pre-trained model path options in scripts.
• Also, you can download our pretrained weight: Link (base session weight)

Citation

@inproceedings{kim2024mics,
  title={MICS: Midpoint Interpolation to Learn Compact and Separated Representations for Few-Shot Class-Incremental Learning},
  author={Kim, Solang and Jeong, Yuho and Park, Joon Sung and Yoon, Sung Whan},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  year={2024}
}