branch-DIPO

Overview

branch-DIPO is a powerful open-source implementation of Density-driven Iterative Prototype Optimization (DIPO) for Transductive Few-shot Learning, designed to streamline research and practical usage.

Features

• Utilizes the same backbone network and training strategies as 'S2M2_R'.
• Pretrained models and feature extractors are available to simplify usage.

Getting Started

Feature Extractors

You can find training details for the backbone network in the S2M2_fewshot repository. Alternatively, download our pretrained features and models from Google Drive and unzip them. Remember to adjust file paths accordingly after extracting the files.

Evaluation

To evaluate the proposed method, run the following command:

python evaluate_DIPO.py

References

branch-DIPO builds upon several impactful research contributions:

• Charting the Right Manifold: Manifold Mixup for Few-shot Learning
• FREE LUNCH FOR FEW-SHOT LEARNING: DISTRIBUTION CALIBRATION