Re:Three open source Sonos projects: efficient embedded deve...
by hobofan on Saturday October 24, 2020 @06:25PM
(#24882385)
Tract supports ONNX models, which you can export from pytorch. See nnsplit[0] as an example for a project that trains with pytorch and serves the model via tract-onnx.
Re:Three open source Sonos projects: efficient embedded deve...
by aldanor on Saturday October 24, 2020 @09:14PM
(#24883125)
You can also export them from LightGBM and XGBoost (see sklearn-onnx) for forests and tract will handle them! // well, almost... I need to finally find some time and finish my tract PR :)
Re:Three open source Sonos projects: efficient embedded deve...
by imglorp on Saturday October 24, 2020 @08:49PM
(#24883029)
We were unable to get Tract to work at $job but there's another project, I think called onnxruntime-rs that we did get working with our models.
Re:Three open source Sonos projects: efficient embedded deve...
by narrationbox on Saturday October 24, 2020 @06:33PM
(#24882437)
Tract looks incredible, I am surprised that it can support Wavenet on the Raspberry Pi Zero.
Re:Three open source Sonos projects: efficient embedded deve...
by woodson on Saturday October 24, 2020 @07:08PM
(#24882617)
I believe it also supports what it calls "pulse" models converted/imported from Kaldi for low-latency ASR/keyword spotting (though this may have changed, I haven't checked in a while).
I'd love if they did this with pytorch. Tensorflow is quickly falling out of favor.
Deploying models in Rust makes a lot of sense. It's a little rough around the edges, though. (I use Rust bindings to libtorch for https://vo.codes )
Sonos has done some scummy things in the past (bricking units to prevent resale or refurbishment, deactivating old hardware, etc), but they operate in a very hard business sector against giants with monopolistic loss leaders, and this is a really cool open source release. I'm not aiming to buy more Sonos equipment, but I like them a little more now.