The KeOps library lets you compute generic reductions of large 2d arrays whose entries are given by a mathematical formula. It is perfectly suited to the computation of convolutions (or more generally to Kernel dot products) and the associated gradients (with an automatic differentiation engine).
KeOps is fast as it allows you to compute Gaussian convolution up to 40 times faster than a standard tensor algebra library that use GPU. KeOps is scalable and can be used on large data (typically from n=10^3 to n=10^7 number of rows/columns): it combines a tiled reduction scheme and works even when the full kernel matrix does not fit into the GPU memory. Finally, KeOps is easy to use as it comes with its Matlab, Python (NumPy or PyTorch) and R (coming soon) bindings.
Web site: http://www.kernel-operations.io
Speaker(s): Dr. Benjamin Charlier,
Location:
Room: ASB 10900
Bldg: Applied Sciences Building
8888 University Drive
School of Engineering Science
Burnaby, British Columbia
V5A 1S6