git clone https://github.com/liangyu-git/FSOR-OPP.git
cd FSOR-OPP
pip install -r requirements.txt

You may need to download datasets inclusive of MiniImageNet and TieredImageNet as well as some pretrained model weights. All of the above vital information can be downloaded from TANE.

cd fsor_dinoEXP # or fsor_resnet12 for a differnet backbone
./run.sh

Arguments:

• --dataset: select the specific dataset to train.
• --logroot: the path for logging.
• --data_root: the path to the dataset.
• --n_ways and --n_shots: select the training methods which is introduced in the paper.
• --restype: adjust the model types.
• pretrained_model_path: the path to the pretrained model weights.
• learning rate: literately, learning rate.
• --gpus: assign which GPU to train the model.
• --n_train_para: select the number of tasks in a training process.
• --n_train_runs: the number of iteration in one epoch.
• op_loss: the alpha rate.
• protonet: if the mode is protonet or not.

If you find this project useful, please consider citing:

@article{sun2024overall,
  title={Overall positive prototype for few-shot open-set recognition},
  author={Sun, Liang-Yu and Chu, Wei-Ta},
  journal={Pattern Recognition},
  volume={151},
  pages={110400},
  year={2024},
  publisher={Elsevier}
}