one hint would be looking into dual numbers for differentiation to make that as painfree as possible, but not necessarily
what ever you want to start with, I'd be pleased to review pull requests
Bernhard Schuster
@drahnr
also don't hesitate to ask questions and discuss stuff
bright-star
@bright-star
Awesome, thanks!
subversive-owl
@subversive-owl
Oh wow there is OpenCL interest
that's actually what i'm here to ask about :)
I have a virtex 5 here I can't get to show up with https://github.com/cogciprocate/ocl 's API, but if I can get it talking, I was interested in trying autumn on it
Bernhard Schuster
@drahnr
@subversive-owl the OpenCL implementation is just being started, there is still way to go
subversive-owl
@subversive-owl
I'm interested in helping! if I can get this FPGA working to test with
Bernhard Schuster
@drahnr
cool :)
I am currently investigating / experimenting on how to get automatic differentiation on board
if you feel like implementing layers that would be awesome :)
subversive-owl
@subversive-owl
sounds good! I haven't done much openCL but I've written parallelized linalg code before, so I can figure it out
Bernhard Schuster
@drahnr
sweet :)
don't hesitate to ask if you are stuck
subversive-owl
@subversive-owl
i'm installing the Xilinx ISE to see if that clears up the FPGA comm issues, then I can get started
Bernhard Schuster
@drahnr
and maybe tickets and PRs as you go so we can discuss on a per topic basis
it will be decoupled from the original branch in the next few weeks
while parenchyma has somewhat similar goals it is very distinct
Bernhard Schuster
@drahnr
for the hadmard product it would probably make sense to create another collenchyma-utils, since it does not really belong into blas http://www.netlib.org/blas/ where I would have put it intuitively
Bernhard Schuster
@drahnr
or you manually setup the hadmard matrix and use blas multiplication operations
subversive-owl
@subversive-owl
I was just about to ask that question :p
the two names on the impressum both announced they were quitting last year, which I just realized
but the publicly stated reason was basically "won't be popular enough with everyone", which doesn't really matter to me
I want to build something fast that runs on metal and is very, very reliable
Aloïs Cochard
@aloiscochard
@drahnr okay, it is very good to hear there is a continuation of the original impl. it looks promising. I did switch to ArrayFire in the meantime but I'll definitely keep an eye on collenchyma and how it evolve, thanks!