Skip to content

Implementation of NNDescent (randomized iterative KNN approximation) optimized for maximal single-core performance.

Notifications You must be signed in to change notification settings

resulknad/nndescent

Repository files navigation

Nearest Neighbour Descent

see report for background: 52_report.pdf

This is the repository for the Advanced Systems Lab project by:

  • Jonas
  • Tobias
  • Dan
  • Samuel

The goal of the project is to implement NN-Descent which is an efficient algorithm for approximate K-Nearest Neighbour Graph (K-NNG) construction.

The algorithm is described in the pulication "Efficient K-Nearest Neighbor Graph Construction for Generic Similarity Measures" by Wei Dong et al.

Useful links:

Installation

After cloning the repository compile the C-code in the directory nn_descent with the following command:

gcc  -O3 -ffast-math -march=native -o a.out knnd.c knnd_test.c vec.c -lm

Additionally you will have to create a Python virtual environment. To do so you can follow these steps:

  1. create the environment in the chosen path (e.g ./pip-env)
    python -m venv ./pip-env
    
  2. activate the environment source pip-env/bin/activate
  3. install the requirements listed in requirements.txt pip install -r requirements.txt

How-To: Benchmark

Obtain code of nn_descent of in the version you wish to benchmark. Usually this means cloning into a new directory and resetting as follows:

git clone ... 
cd copy_of_project
git reset --hard T0

Don't forget to adjust the frequency constant in the C code.

Run

python benchmark.py -p ../copy_of_project/nn_descent

which will generate a tab separated file in directory benchmarks. Make sure to add an identifier of your architectures to files you wish to push to the repistory.

Plotting

You may find examples of plots in eval_zenv1.ipynb

About

Implementation of NNDescent (randomized iterative KNN approximation) optimized for maximal single-core performance.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published