A Maximum Log-Likelihood Method for Imbalanced Few-Shot Learning Tasks

Introduction

This repository provides the algorithm demonstration for the paper A Maximum Log-Likelihood Method for Imbalanced Few-Shot Learning Tasks.

Getting Started

Install dependencies (I prefer a conda environment)

• conda create -n MLL_FSL python=3.8.13
• conda activate MLL_FSL
• pip install -r requirements.txt
• python -m ipykernel install --user --name MLL_FSL
• python setup.py develop (or python setup.py install if you don't want to do development)

Download preprocessed features and trained models

• This data is over 20GB and we are currently in the process of finding a public drive to host the data.

Reproducing the results

• cd ./scripts/
• bash run_all_inductive.sh to run all inductive results
• bash run_all_transductive.sh to run all transductive results

Contact

For further questions or details, reach out to Samuel Hess (shess@email.arizona.edu)

Acknowledgements

Special thanks to the authors of many prior works that have shared their code, including:

• SimpleShot

• FEAT

• TIM

• αTIM