Classification Extension

This document describes Triton's classification extension. The classification extension allows Triton to return an output as a classification index and (optional) label instead of returning the output as raw tensor data. Because this extension is supported, Triton reports “classification” in the extensions field of its Server Metadata.

An inference request can use the “classification” parameter to request that one or more classifications be returned for an output. For such an output the returned tensor will not be the shape and type produced by the model, but will instead be type BYTES with shape [ batch-size, <count> ] where each element returns the classification index and label as a single string. The <count> dimension of the returned tensor will equal the “count” value specified in the classification parameter.

When the classification parameter is used, Triton will determine the top-n classifications as the n highest-valued elements in the output tensor compared using the output tensor’s data type. For example, if an output tensor is [ 1, 5, 10, 4 ], the highest-valued element is 10 (index 2), followed by 5 (index 1), followed by 4 (index 3), followed by 1 (index 0). So, for example, the top-2 classifications by index are [ 2, 1 ].

The format of the returned string will be “<value>:<index>[:<label>]”, where <index> is the index of the class in the model output tensor, <value> is the value associated with that index in the model output, and the <label> associated with that index is optional. For example, continuing the example from above, the returned tensor will be [ “10:2”, “5:1” ]. If the model has labels associated with those indices, the returned tensor will be [ “10:2:apple”, “5:1:pickle” ].

HTTP/REST

In all JSON schemas shown in this document $number, $string, $boolean, $object and $array refer to the fundamental JSON types. #optional indicates an optional JSON field.

The classification extension requires that the “classification” parameter, when applied to a requested inference output, be recognized by Triton as follows:

• “classification” : $number indicating the number of classes that should be returned for the output.

The following example shows how the classification parameter is used in an inference request.

POST /v2/models/mymodel/infer HTTP/1.1
Host: localhost:8000
Content-Type: application/json
Content-Length: <xx>
{
  "id" : "42",
  "inputs" : [
    {
      "name" : "input0",
      "shape" : [ 2, 2 ],
      "datatype" : "UINT32",
      "data" : [ 1, 2, 3, 4 ]
    }
  ],
  "outputs" : [
    {
      "name" : "output0",
      "parameters" : { "classification" : 2 }
    }
  ]
}

For the above request Triton will return the “output0” output tensor as a STRING tensor with shape [ 2 ]. Assuming the model produces output0 tensor [ 1.1, 3.3, 0.5, 2.4 ] from the above inputs, the response will be the following.

HTTP/1.1 200 OK
Content-Type: application/json
Content-Length: <yy>
{
  "id" : "42"
  "outputs" : [
    {
      "name" : "output0",
      "shape" : [ 2 ],
      "datatype"  : "STRING",
      "data" : [ "3.3:1", "2.4:3" ]
    }
  ]
}

If the model has labels associated with each classification index Triton will return those as well, as shown below.

HTTP/1.1 200 OK
Content-Type: application/json
Content-Length: <yy>
{
  "id" : "42"
  "outputs" : [
    {
      "name" : "output0",
      "shape" : [ 2 ],
      "datatype"  : "STRING",
      "data" : [ "3.3:1:index_1_label", "2.4:3:index_3_label" ]
    }
  ]
}

GRPC

The classification extension requires that the “classification” parameter, when applied to a requested inference output, be recognized by Triton as follows:

• “classification” : int64_param indicating the number of classes that should be returned for the output.

The following example shows how the classification parameter is used in an inference request.

ModelInferRequest {
  model_name : "mymodel"
  model_version : -1
  inputs [
    {
      name : "input0"
      shape : [ 2, 2 ]
      datatype : "UINT32"
      contents { int_contents : [ 1, 2, 3, 4 ] }
    }
  ]
  outputs [
    {
      name : "output0"
      parameters [
        {
          key : "classification"
          value : { int64_param : 2 }
        }
      ]
    }
  ]
}

For the above request Triton will return the “output0” output tensor as a STRING tensor with shape [ 2 ]. Assuming the model produces output0 tensor [ 1.1, 3.3, 0.5, 2.4 ] from the above inputs, the response will be the following.

ModelInferResponse {
  model_name : "mymodel"
  outputs [
    {
      name : "output0"
      shape : [ 2 ]
      datatype  : "STRING"
      contents { bytes_contents : [ "3.3:1", "2.4:3" ] }
    }
  ]
}