Repository navigation
Convolutions and Pooling #102
Description
See here for a description of the operations involved.
I expect that pooling will be relatively easy, convolutions are also straightforward enough in principle but hard to make really fast. We probably want to port or wrap an existing high-quality implementation for Open CL, but I'm not sure what the benchmarks is here (perhaps DeepCL?).
Activity
tejank10 commentedon Feb 23, 2018
I am interested to take this up! From what I interpret, do we require the GPUArray version of the convolutions and pooling in Julia? Please correct me if I am wrong.
MikeInnes commentedon Mar 7, 2018
Yes, that or some wrapper for an existing implementation. A generic kernel is probably a good idea either way though.
XVilka commentedon Dec 25, 2019
@tejank10 have you had any luck with this?
maleadt commentedon Oct 19, 2026
Convolutions and pooling are provided by NNlib.jl, with generic GPU kernels and vendor-library back-ends (cuDNN, MIOpen), so this does not belong in GPUArrays.