KeOps: Kernel operations on the GPU, with autodiff, without memory overflows

Paul Escande
I2M, CNRS, Marseille
/user/paul.escande/

Date(s) : 23/04/2021   iCal
14 h 30 min - 15 h 30 min

The KeOps library is developed by Benjamin Charlier, Jean Feydy and Joan Alexis Glaunès.

It lets you compute reductions of large arrays whose entries are given by a mathematical formula or a neural network. It combines efficient C++ routines with an automatic differentiation engine and can be used with Python (NumPy, PyTorch), Matlab and R.
It is perfectly suited to the computation of kernel matrix-vector products, K-nearest neighbors queries, N-body interactions, point cloud convolutions and the associated gradients. Crucially, it performs well even when the corresponding kernel or distance matrices do not fit into the RAM or GPU memory. Compared with a PyTorch GPU baseline, KeOps provides a x10-x100 speed-up on a wide range of geometric applications, from kernel methods to geometric deep learning.

DISCLAIMER: I am NOT one of the authors of the library, I happen to use it and thought it could be of interest for many researchers. My knowledge of the library is thus not complete but I believe this seminar would provide a glimpse of its operation.

https://www.kernel-operations.io/keops/index.html

Catégories



Retour en haut