[ICLR2021 Oral] Free Lunch for Few-Shot Learning: Distribution Calibration
paper link: https://openreview.net/forum?id=JWOiYxMG92s
zhihu link: https://zhuanlan.zhihu.com/p/344531704
Backbone Training
We use the same backbone network and training strategies as 'S2M2_R'. Please refer to https://github.com/nupurkmr9/S2M2_fewshot for the backbone training.
Extract and save features
After training the backbone as 'S2M2_R', extract features as below:
-
Create an empty 'checkpoints' directory.
-
Run:
python save_plk.py --dataset [miniImagenet/CUB]
Or you can directly download the extracted features/pretrained models from the link:
https://drive.google.com/drive/folders/1IjqOYLRH0OwkMZo8Tp4EG02ltDppi61n?usp=sharing
After downloading the extracted features, please adjust your file path according to the code.
Evaluate our distribution calibration
To evaluate our distribution calibration method, run:
python evaluate_DC.py
Citation
If our paper is useful for your research, please cite our paper:
@inproceedings{
yang2021free,
title={Free Lunch for Few-shot Learning: Distribution Calibration},
author={Yang, Shuo and Liu, Lu and Xu, Min},
booktitle={International Conference on Learning Representations (ICLR)},
year={2021},
}
Reference
Charting the Right Manifold: Manifold Mixup for Few-shot Learning