Availability of All-In-One-Small?

[jonahsiegle]

Hey there,

First of all thank you for your work on this model! The accuracy of beat detection and segment labeling is the best I've ever seen.
Over the past few months I've been working on an iOS port for portions of this model. I have things working and running predictions. However the obvious bottleneck is memory usage on mobile (A one minute track chunk uses over 2GB of memory for example).

I saw in the paper you mentioned a 46K parameter smaller version of the model that has similar results to the regular one. I assume the smaller one generally uses less memory. So my question is how does one go about using the small version? Is this something trivial for me to tweak in the current code base or is this an entirely different codebase/architecture and training checkpoints?

Any guidance would be amazing, I'm happy to help contribute if this is something to tweak in the current codebase.