 |
RKeOps v2: Kernel operations with Symbolic Tensors on the GPU in R
1 : Institut Montpelliérain Alexander Grothendieck
Centre National de la Recherche Scientifique, Université de Montpellier
2 : Laboratoire de biologie et modélisation de la cellule
* : Auteur correspondant
École Normale Supérieure - Lyon, Université Claude Bernard Lyon 1, Institut National de la Santé et de la Recherche Médicale : U1210, Centre National de la Recherche Scientifique : UMR5239, Institut National de la Santé et de la Recherche Médicale, Centre National de la Recherche Scientifique
We present RKeOps version 2, a binder in R for the KeOps library which implements "kernel operations on the GPU, with autodiff, without memory overflows". The RKeOps package allows the user to seamlessly perform fast and memory-efficient kernel computations, based on symbolic matrix operations, such as kernel matrix reductions, convolutions or nearest neighbor search, involving large datasets (up to 1E7 points), on CPU or GPU without any additional development cost. The main contribution of this work is to provide the LazyTensor abstraction directly in R, allowing to write tensor operations similarly to R native syntax for vector and matrix operations, greatly simplifying the user experience.
