One-Shot Domain Incremental Learning

This is an official Pytorch implementation of "One-Shot Domain Incremental Learning". This study was presented at IJCNN 2024. [arXiv|IEEE Xplore]

Preparation

1. Git clone this repository.

   git clone https://github.com/ToyotaCRDL/OneShotDIL.git

2. Build a Docker image using Dockerfile and requrements.txt.

   docker build --build-arg USER=${USER} -t one-shot-dil OneShotDIL

3. Launch a container.

   docker run -it --rm --gpus all \
           --mount type=bind,source=$(pwd)/OneShotDIL,target=${HOME}/OneShotDIL \
           --workdir ${HOME}/OneShotDIL \
           one-shot-dil

We conducted our experiments on a NVIDIA A100 GPU.

Training

• The following is a demonstration of CIFAR10.

  1. Train a model using only the original domain.

     bash scripts/cifar10/only_original_domain.sh

  2. Perform one-shot domain incremental learning using the new domain.

     • Standard method (updated-stats)

       bash scripts/cifar10/updated_stats_dil.sh

     • Proposed method (fixed-stats)

       bash scripts/cifar10/fixed_stats_dil.sh

Evaluation

• Calculate the test accuracy before and after performing one-shot domain incremental learning.

  • Standard method (updated-stats)

    bash scripts/cifar10/updated_stats_eval.sh

  • Proposed method (fixed-stats)

    bash scripts/cifar10/fixed_stats_eval.sh

Citation

To cite our work, you can use the following:

@inproceedings{OneShotDIL,
  author={Esaki, Yasushi and Koide, Satoshi and Kutsuna, Takuro},
  booktitle={2024 International Joint Conference on Neural Networks (IJCNN)}, 
  title={One-Shot Domain Incremental Learning}, 
  year={2024},
  doi={10.1109/IJCNN60899.2024.10650928}
}

License

Copyright (C) 2025 TOYOTA CENTRAL R&D LABS., INC. All Rights Reserved.