puffinn / esa-paper

Supplemental material for the ESA 2019 paper describing PUFFINN

Purpose

This repository contains supplemental material to the ESA 2019 paper "PUFFINN: Parameterless and Universally Fast FInding of Nearest Neighbors". This repository contains the scripts and Jupyter notebooks used in the evaluation. The code is available at https://github.com/puffinn/puffinn. Please refer to this repository to build the code.

Replication of Experiments in Paper

1. Get ann-benchmarks from https://github.com/erikbern/ann-benchmarks
2. Copy algos-puffinn-evaluation.yaml into the main ann-benchmarks directory.
3. Copy Dockerfile.puffinn into ann_benchmarks/install/.
4. Install ann-benchmarks following the guide on https://github.com/erikbern/ann-benchmarks
5. Run python run.py --dataset glove-100-angular --definitions algos-puffinn-evaluation.yaml from the main ann-benchmarks directory to run all experiments on Glove-1M. (Note that this will take several days. Modify the list of parameters used by PUFFINN to shorten the time it takes to carry out the experiment.)

Looking at the Evaluation

All raw experimental results are found in evaluation/all.csv. Data and plots were evaluated using Jupyter notebook. The notebook is available at evaluation/log.ipynb.