Instead of Nvidia, can we use OpenVino for Intel Processors?

[Dvalin21]

Same as the title, and if so, please share how. Thanks.

[pospielov]

CompreFace doesn't support OpenVino now.
The problem is that to add support for CUDA, all we need is to install tensorflow-gpu library instead of tensorflow.
OpenVino is another framework. I played around with it a couple of years ago and it was very raw without proper documentation.
I hope Intel will add somehow support for OpenVino to popular libraries like tensorflow and pytorch

[Dvalin21]

Thanks!

[PovilasID]

Frigate NVR now has an option to run on OpenVINO and it's awesome (low power consumption and fast). I would love to be able to reduce your power consumption or not to have add some crappy GPU to the system just for to run this...

It looks like OpenVINO now has tensorflow integration https://docs.openvino.ai/latest/ovtf_integration.html and model conversion option for pytorch https://docs.openvino.ai/latest/openvino_docs_MO_DG_prepare_model_convert_model_Convert_Model_From_PyTorch.html

Though docs for both look... limited. Not sure if that is what you meant "without proper documentation" but if you could take a second look would be very nice

[pospielov]

Here is a trick. If you want to use OpenVino, you can use lots of converters (and yes, when I tried it several years ago, there did not exist).
But it still means we need to create a separate build for compreface-core that uses OpenVino libraries, and put there converted models.
The ideal situation for us will be if we continue to use tensorflow without changing the code, as it works for nvidia GPUs now.
We discovered that ONNXRuntime gives such an ability, but we will still need to migrate compreface-core to it.

[Dvalin21]

I look forward to the updates