[ICLR2021 Oral] Free Lunch for Few-Shot Learning: Distribution Calibration

paper link: https://openreview.net/forum?id=JWOiYxMG92s

Requirements

To install requirements:

pip install -r requirements.txt

Download the dataset and create base/val/novel splits:

CUB

Change directory to filelists/CUB/
Run 'source ./download_CUB.sh'

Train feature extractor

To train the feature extractor in the paper, run this command:

python train.py --dataset [miniImagenet/CUB]

Extract and save features

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

Actually, all our algorithm is built upon the extracted features (We perform data augmentation in the feature space). The training procedure and the pretrained backbone are both irrelevant to our method. The pretrained model and the extracted features we used are the same as the reference work 'S2M2' (Their project page https://github.com/nupurkmr9/S2M2_fewshot). You can reproduce our work by just simply applying evaluate_DC.py on the provided features. Or you can apply our method on your own model.

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

Reference

Charting the Right Manifold: Manifold Mixup for Few-shot Learning

https://github.com/nupurkmr9/S2M2_fewshot

TheStarkor/fewshotDC