Add metadata to pipeline outputs

[xiaofei-du]

Our VDP protocol is flexible enough to add metadata to the outputs.

For example: adding latency field

[
  {
    "latency": 11,   -----> pipeline-backend <-> model-backend latency in ms
    "task": "CLASSIFICATION",
    "batch_outputs": [
      {
         "classification": 
         {
             "category": "dog",
             "score": 0.8
         },
         "latency": 10  -----> model-backend <-> triton latency in ms
      }
    ]
  },
  {
     "task": "DETECTION",
     ...
  }
]

[pinglin]

Considering to add a metadata object peering with the classification object for any general attributes for each batch image.

[pinglin]

Below is a concrete proposal for pipeline data mapping.

For SYNC pipeline triggering response:

{
  "data_mapping_indices": [
    "input1-uuid",
    "input2-uuid",
    ...
  ],
  "model_instance_outputs": [
    {
      "model_instance": "models/mobilenetv2/instances/v1.0"
      "task": "TASK_CLASSIFICATION",
      "batch_outputs": [
        {
          "index": "input1-uuid",
          "classification": {
            "category": "golden retriever",
            "score": 0.898938
          }
        },
        {
          "index": "input2-uuid",
          "classification": {
            "category": "ice bear",
            "score": 0.999554
          }
        },
        ....
      ]
    }
  ]
} 

and for ASYNC triggering response:

{
  "data_mapping_indices": [
    "input1-uuid",
    "input2-uuid",
    ...
  ],
  "model_instance_outputs": [
    {
      "model_instance": "models/mobilenetv2/instances/v1.0"
      "task": "TASK_CLASSIFICATION",
      "batch_outputs": [
        {
          "index": "input1-uuid",
          "classification": null
        },
        {
          "index": "input2-uuid",
          "classification": null
        },
        ....
      ]
    }
  ]
}

The model-backend triggering response can be kept as the same since from the model-backend’s point of view, there is nothing to map.