If you’ve regarded into GPU-accelerated machine studying tasks, you’re actually accustomed to NVIDIA’s CUDA structure. It additionally follows that you just’ve checked the costs on-line, and know the way costly it may be to get a high-performance video card that helps this specific model of parallel programming.
But what if you happen to might run machine studying duties on a GPU utilizing nothing extra unique than OpenGL? That’s what [lnstadrum] has been engaged on for a while now, as it will enable gadgets as meager as the unique Raspberry Pi Zero to run duties like picture classification far quicker than they may utilizing their CPU alone. The trick is to interrupt down your computational job into one thing that may be carried out utilizing OpenGL shaders, that are typically meant to push online game graphics.
An instance of X2’s neural web upscaling.
[lnstadrum] explains that OpenGL releases from the final decade or so truly embrace so-called compute shaders particularly for operating arbitrary code. But sadly that’s not an choice on boards just like the Pi Zero, which solely meets the OpenGL for Embedded Systems (GLES) 2.0 normal from 2007.
Constructing the neural web in such a manner that it will be appropriate with these extra constrained platforms was far more troublesome, however the finish outcome has way more attention-grabbing functions to point out for it. During exams, each the Raspberry Pi Zero and a number of other older Android smartphones had been in a position to run a pre-trained picture classification mannequin at a good fee.
This isn’t just a few thought experiment, [lnstadrum] has launched a picture processing framework known as Beatmup utilizing these ideas that you may mess around with proper now. The C++ library has Java and Python bindings, and in keeping with the documentation, ought to run on just about something. Included within the framework is an easy instrument known as X2 which might carry out AI picture upscaling on every little thing out of your laptop computer’s built-in video card to the Raspberry Pi; making it a good way to take a look at this fascinating software of machine studying.
Truth be instructed, we’re a bit behind the ball on this one, as Beatmup made its first public launch again in April of this yr. It might need flown beneath the radar till now, however we predict there’s lots of potential for this mission, and hope to see extra of it as soon as phrase will get out concerning the spectacular outcomes it might probably wring out of even the lowliest {hardware}.
[Thanks to Ishan for the tip.]
OpenGL Machine Learning Runs On Low-End Hardware
