ideas..
toy library, implementation of convnet inference . got a vague goal of upscaling retro games.. a simple lib that could be integrated into emulators. Probably better the use ONNX or whatever but I want to tinker with low level ideas and raw maths first.
It's a kind of warm up and diversion , I'm mainly focussed on a rust game engine. I'd like to use NN's to enhance or generate content there.
unlikely to implement training here - train in PyTorch? make a simplified export format to read in?
various concerns -
Rust vs C++ - I use use mostly these days but have gone back to c++ for this (useful to keep C++ ability up and its still way more popular. ).
use of OpenCL - am working on an M1 Mac mini, in the end there are capabilities unavailable eg frustratingly Apple's deprecated OpenCL wont do fp16, and it can't use tensor cores for Nvidia (?) .. might have to bite the bullet and make multiple back ends (at which point a pro lib would be better)
ideas.. toy library, implementation of convnet inference . got a vague goal of upscaling retro games.. a simple lib that could be integrated into emulators. Probably better the use ONNX or whatever but I want to tinker with low level ideas and raw maths first. It's a kind of warm up and diversion , I'm mainly focussed on a rust game engine. I'd like to use NN's to enhance or generate content there. unlikely to implement training here - train in PyTorch? make a simplified export format to read in?
various concerns - Rust vs C++ - I use use mostly these days but have gone back to c++ for this (useful to keep C++ ability up and its still way more popular. ). use of OpenCL - am working on an M1 Mac mini, in the end there are capabilities unavailable eg frustratingly Apple's deprecated OpenCL wont do fp16, and it can't use tensor cores for Nvidia (?) .. might have to bite the bullet and make multiple back ends (at which point a pro lib would be better)