Benchmarks

These benchmarks showcase the performances of the KeOps routines as the number of samples/points varies (typical use cases should be from 100 to 1,000,000).

Comparison with other related projects

KeOps is fast! You may find here a benchmark that compare the performances of pyKeOps with other related projects.

pyKeOps benchmarks

Note

If you run a KeOps script for the first time, the internal engine may take a few minutes to compile all the relevant formulas. This work is done once as KeOps stores the resulting shared object files (*.so) in a cache directory.

Utility functions for the benchmarks

Datasets for the benchmarks

Mixed-precision and accuracy settings

K-Nearest Neighbors search

Radial kernels convolutions

Gradient of Radial kernels convolutions

Benchmarking Gaussian convolutions in high dimensions

Solving positive definite linear systems

Scaling up multi-head attention layers

Scaling up Gaussian convolutions on 3D point clouds