Google Coral Edge TPU support?

[Alphakilo]

Hi there 👋

I wanted to ask whether there's the possibility to use a Google Coral TPU to speed up classification?

Thanks!

[muebau]

I think it could be helpful to have some kind of interface like suggested in #680 . It should be easy to comply to such interface with some docker container.
As there is WebUSB support now (https://github.com/google-coral/webcoral) it should be easy to create a browser version like I suggested in #680 too.

[Ayo91]

amazing idea!

[Dummay]

As Tensorflow is used by Nextcloud it should be fairly simple to replicate how other projects like Frigate NVR leverage Google Coral TPU, see here https://docs.frigate.video/frigate/hardware, at least that's what I in my infinite ignorance think :)

[marcelklehr]

The Coral Edge TPU is capable of executing deep feed-forward neural networks such as convolutional neural networks (CNN). It supports only TensorFlow Lite models that are fully 8-bit quantized and then compiled specifically for the Edge TPU.

Recgonize does not exclusively use TF lite models. Anything that works in WASM mode probably also works in TF lite, but it would be an additional effort to compile models for TF lite and add support for coral tpus.

[Ayo91]

So, Coral TPUs are not supported in Nextcloud, which gpus do you recommend for speed up the recognition processes?