Running inference inside a Bela?

markh

Hi,

I'm considering buying a Bela for running SuperCollider, with saxophone as the main input to interactive patches written in SC, sort of like an intelligent effects pedal.

What I'm wondering is, do you know if Bela's hardware has capacity to also run a pretrained neural network alongside SC or would I be better off running that on a separate device? It'd communicate via OSC with SuperCollider. I was thinking to move the neural net into a more performative C++ script so that it'd run smoother (it's in Python at the moment).

Does this sound feasible to anyone on here? I'd love to hear of any examples of people trying this sort of thing.

Cheers,

Mark

adanlbenito

Hi Mark,

It should definitely be possible to run a neural net in Bela. Something along those lines has already been done for this paper that will be presented at ICASSP next week.

Writing your own C++ implementation is probably the easiest way to get inference running within Bela, although TensorFlow Lite should also be able to run on the BealeBone black (however this has not been tested with Bela to the best of my knowledge).

If you want to run inference alongside a complex patch, you would however need to be careful with how you manage your threads and keep in mind that there is only so much you can do within the limits of the CPU.

markh

Hi, thanks so much for the reply! I missed it at the time. Great to know that it's at least possible.

Exceeding CPU limit in the Bela was definitely one of my concerns - I don't think my SuperCollider patches are especially hungry, but then I've only tested them on my laptop.

My preference would very much be running a TensorFlow Lite model over writing one in C++, perhaps we don't share the same concept of 'easiest'! 😃

Exciting! I will let the forum know how I get on,

Thanks again

Mark