terkaslan/lmi

Accompanying source code for Learned Metric Index — Proposition of learned indexing for unstructured data

This repo has become obsolete in favor of a more user-friedly implementation of LMI -- TerkaSlan/Learned-Metric-Index-Framework

Learned-indexes

Paper accompanying this code

@article{antol2021learned,
  title={Learned Metric Index—Proposition of learned indexing for unstructured data},
  author={Antol, Matej and Ol’ha, Jaroslav and Slanin{\'a}kov{\'a}, Ter{\'e}zia and Dohnal, Vlastislav},
  journal={Information Systems},
  volume={100},
  pages={101774},
  year={2021},
  publisher={Elsevier}
}

Link to the Elsevier's Information Systems issue will be provided once the journal is published.

Intro

This repository contains the code necessary to build with an example CoPhIR dataset of 100k objects included. Currently supported ML models:

• Fully-connected NNs
• Logistic regression
• Fully-connected NNs trained in a Multi-label fashion
• Random Forests

How to run

• Download the repo and run the code locally
  • python 3.x required
  • run pip install -r requirements.txt to install needed dependencies
  • run the introductory notebook (LMI Playground.ipynb), optionally download the dataset using download_data.sh