skingorz/FD-Align

FD-Align: Feature Discrimination Alignment for Fine-tuning Pre-Trained Models in Few-Shot Learning

Officaial implementation of FD-Align for paper "FD-Align: Feature Discrimination Alignment for Fine-tuning Pre-Trained Models in Few-Shot Learning" (NeurIPS 2023)

This code is based on LightningSL. You can refer to corss modal adaptation and Channel_Importance_FSL to prepare the data.

CoOp task

1. Fine-tune the classification head (Optional).

bash tools/LP_alldataset.sh

2. FD-Align Fine-tune

bash tools/FT_alldataset.sh

N-way-K-shot task

Train

Fine tune model on miniImageNet.

bash tools/train_meta.sh

Test

Evaluate performance on different datasets.

bash tools/test_meta.sh

Citation

If our code is helpful for your research, please cite the following paper:

@article{song2023FD,
    title={FD-Align: Feature Discrimination Alignment for Fine-tuning Pre-Trained Models in Few-Shot Learning},
    author={Kun Song and Huimin Ma and Bochao Zou and Huishuai Zhang and Weiran Huang},
    journal={NeurIPS},
    year={2023}
}