Custom Vision model exported to Docker file and Vision AI Developer Kit.

Copper Contributor

Hi,

I have a camera (Vision AI DevKit) and it's already registered as an IoT Edge Device on Azure, its two modules are running (AIVisionDevKitGetStartedModule and WebStreamModule). Then I created and trained a model using customvision.ai, exported this trained model to Dockerfile (not to Vision Ai DevKit).

 

Is there a tutorial that I can follow to take this Dockerfile and deploy it into the Vision AI DevKit?

 

Cheers!

4 Replies

Dear @PuneetRaheja 

 

I have previously followed the https://azure.github.io/Vision-AI-DevKit-Pages/docs/Tutorial-HOL_Using_the_VisionSample/ and I can see the results on the camera, although that one uses as "Export Capabilities" the Vision AI Dev Kit. On customvision.ai, if I set the model to Vision AI Dev Kit, I can only run a "quick training" and I usually get a very low Recall; if instead of using Vision AI Dev Kit on the Export Capabilities I use "Basic platforms (Tensorflow, CoreML, ONNX, ...)" then I can run an advanced training and I get a considerable higher Recall. That is the reason why I created my own model on customvision.ai, I trained it with an advanced training and exported it to Dockerfile...but now I can't find the right way to deploy it to the Vision AI Dev Kit on Azure IoT Edge since it seems to take the model exclusively from the AIVisionDevKitGetStartedModule.

 

 

@Guipsot you can’t use the docker image to deploy.

VAIDK has its proprietary Qualcomm AI hardware accelerator (SNPE).

The VAIDK export file DLC supports SNPE but the docker file one doesn’t.

Custom Vision only supports quick training for Vision AI Dev kit now.

@Guipsot 

 

If you want to deploy an AI model from Customvision.ai easily I'd recommend using module twin update that is described in the tutorials. 

 

In case you want to build your own container around it you'll find examples from Github: https://github.com/microsoft/vision-ai-developer-kit/tree/master/samples/research/VisionSample