[Reproducability Challenge 2021] Free Lunch for Few-Shot Learning: Distribution Calibration

FREE LUNCH FOR FEW-SHOT LEARNING: Distribution Calibration written by Shuo Yang, Lu Liu, Min Xu is to transfer statistics from base classes with enough examples to calibrate the distribution of these few-sample classes, and then to draw a sufficient number of examples from the calibrated distribution to expand the input of the classifier. The calibrated distribution is then drawn from a sufficient number of examples to expand the input to the classifier Yang et al. (2021). By running the Distribution Calibration code in the appendix of this paper

and pre-training the data, we will confirm whether the results mitigate the overfit- ting phenomenon in few-sample learning, as claimed in this paper. By calculating

the accuracy of SVM and logistic regression, Tukey transformation and the pres- ence or absence of generated features, we see that Distribution Calibration does

have some improvement on the overfitting problem.

Original paper

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

Requirements

• numpy==1.17.2
• matplotlib==3.1.1
• tqdm==4.36.1
• torchvision==0.6.0
• torch==1.5.0
• Pillow==7.1.2

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

Team members

• Yijun Chen [email protected]
• Shuning Ling [email protected]
• Pasinpat Vitoochuleechoti [email protected]

Original Code

https://github.com/ShuoYang-1998/Few_Shot_Distribution_Calibration