Upgrade OpenVINO to 2023.1.0

[felipecrs]

This is just a reminder that OpenVINO has a new version, and I suppose Frigate should eventually upgrade:

• https://www.phoronix.com/news/OpenVINO-2023.1-Released
• https://github.com/openvinotoolkit/openvino/releases/tag/2023.1.0

[NickM-27]

The problem with this is that the NCS2 is no longer supported with this version, so we would be losing support for some devices currently being used (#7319)

[felipecrs]

Damn. Nothing is simple in this world. 😅

[svenstaro]

There's now 2023.3.0 and I think at some point frigate should just upgrade to enable newer opsets. Keeping compatibility with the NCS2 at all costs doesn't seem worthwhile to me.

[jsapede]

upgrade is needed as all new models includes opset 11

[klingon888]

upgrade is needed as all new models includes opset 11

Just saw this post and totally agree that upgrade is more important than keeping compatibility. Since yolov8 is no longer supported, I managed to compile a new yolox-tiny today but it failed miserably when starting Frigate since it has opset11. I assume this error is caused by the old Openvino?

"Cannot create Interpolate layer /backbone/upsample/Resize id:303 from unsupported opset: opset11"

[jsapede]

Could it be done using same steategy as immich where hardware accélération is splitted in docker-compose settings using différent docker images as source for the ai engine ?
Would it make it possible to so have old openVINO for compatibility and newer ones for yolov8/v9 ?

[jsapede]

upgrade is needed as all new models includes opset 11

Just saw this post and totally agree that upgrade is more important than keeping compatibility. Since yolov8 is no longer supported, I compiled a new yolox-tiny today and it failed miserably since it has opset11. I assume this error is caused by the old Openvino?

"Cannot create Interpolate layer /backbone/upsample/Resize id:303 from unsupported opset: opset11"

Theres Somewhere in this github discussion on yolo with a link to a Google collab for yolov8 pretrained for frigate. It stills work with old opset settings

But it's impossible to use the openVINO notebooks for tuning v8 nor v9 : opset error

[klingon888]

Theres Somewhere in this github discussion on yolo with a link to a Google collab for yolov8 pretrained for frigate. It stills work with old opset settings

But it's impossible to use the openVINO notebooks for tuning v8 nor v9 : opset error

I've been using this collab too in 0.13 versions. BUT unfortunately there is an ongoing AGPL issue with using yolov models, so no longer supported in 0.14 versions of Frigate. See here. Only yolo tiny is officially supported now. If you enter yolov8 in the 0.14 config, it now errors out.

[jsapede]

This sucks but in fact for the use we have theres no real différence in models résults, almost all Can acurately distinguish a human from a car or a dog.
The only Real point is the openVINO hardware accélération