Analyze video using custom Yolov3 - Azure Video Analyzer

Question

Hi,
I want to deploy a custom yolov3 to Azure video analyzer following this AVA tutorial analyze-live-video-use-your-model-http. However, it couldn't find the inference server http://yolov3/score.

Instead of using the pre-built docker yolov3 image (mcr.microsoft.com/ava-utilities/avaextension:http-yolov3-onnx-v1.0), I am using a custom docker yolov3 image (deployed on Microsoft container registry). I follow this document http-cpu to push my custom docker image to the container registry.

Now, the Yolov3 has been successfully deployed on "$CONTAINER_REGISTRY_USERNAME_myacr.azurecr.io".
I update the deployment.yolov3.template.json and point the yolov3 module to the mcr image url as follows:

"yolov3": {       
      .....       
     "settings": {       
              "image": "$CONTAINER_REGISTRY_USERNAME_myacr.azurecr.io/avaextension:latest",       
              "createOptions": {}       
 }       

Note that I follow all steps exactly described in AVA tutorial analyze-live-video-use-your-model-http. The only step I'm doing different is to modify the docker image url in the deployment.yolov3.template.json.

When activating the live pipeline, it couldn't find the inference server:

 "body": {   
    "code": "nameResolutionError",   
    "target": "http://yolov3/score"   
  },   
  "properties": {   
    "topic": "/subscriptions/f7fdb0ad-2bed-46f9-b58c-dddec7b9a603/resourceGroups/ava-sample-99/providers/Microsoft.Media/videoAnalyzers/avasampleqsce3xyv5x6wu",   
    "subject": "/edgeModules/avaedge/livePipelines/Sample-Pipeline-1/processors/httpExtension",   
    "eventType": "Microsoft.VideoAnalyzer.Diagnostics.NetworkError",   
    "eventTime": "2022-02-18T07:27:48.295Z",   
    "dataVersion": "1.0"   
  }   

I doubt that the inference url (in operation.json) needs to be modified, but I don't know how. Here is where we set the live pipeline:

{       
    "opName": "livePipelineSet",       
    "opParams": {       
        "name": "Sample-Pipeline-1",       
        "properties": {       
            "topologyName": "EVROnMotionPlusHttpExtension",       
            "parameters": [       
                {       
                    "name": "httpAIServerAddress",       
                    "value": "http://yolov3/score"       
                },...       
            ]       
        }       
    }       
}       

Could you please help me identify the issues? Thanks.

Answer 1

Hi, the docker image takes time to start. After 10 seconds, the model starts and successfully gives inference.