RESTful API

In addition to gRPC APIs TensorFlow ModelServer also supports RESTful APIs. This page describes these API endpoints and an end-to-end example on usage.

The request and response is a JSON object. The composition of this object depends on the request type or verb. See the API specific sections below for details.

In case of error, all APIs will return a JSON object in the response body with error as key and the error message as the value:

{
  "error": <error message string>
}

Model status API

This API closely follows the ModelService.GetModelStatus gRPC API. It returns the status of a model in the ModelServer.

URL

GET http://host:port/v1/models/${MODEL_NAME}[/versions/${VERSION}|/labels/${LABEL}]

Including /versions/${VERSION} or /labels/${LABEL} is optional. If omitted status for all versions is returned in the response.

Response format

If successful, returns a JSON representation of GetModelStatusResponse protobuf.

Model Metadata API

This API closely follows the PredictionService.GetModelMetadata gRPC API. It returns the metadata of a model in the ModelServer.

URL

GET http://host:port/v1/models/${MODEL_NAME}[/versions/${VERSION}|/labels/${LABEL}]/metadata

Including /versions/${VERSION} or /labels/${LABEL} is optional. If omitted the model metadata for the latest version is returned in the response.

Response format

If successful, returns a JSON representation of GetModelMetadataResponse protobuf.

Classify and Regress API

This API closely follows the Classify and Regress methods of PredictionService gRPC API.

URL

POST http://host:port/v1/models/${MODEL_NAME}[/versions/${VERSION}|/labels/${LABEL}]:(classify|regress)

Including /versions/${VERSION} or /labels/${LABEL} is optional. If omitted the latest version is used.

Request format

The request body for the classify and regress APIs must be a JSON object formatted as follows:

{
  // Optional: serving signature to use.
  // If unspecifed default serving signature is used.
  "signature_name": <string>,

  // Optional: Common context shared by all examples.
  // Features that appear here MUST NOT appear in examples (below).
  "context": {
    "<feature_name3>": <value>|<list>
    "<feature_name4>": <value>|<list>
  },

  // List of Example objects
  "examples": [
    {
      // Example 1
      "<feature_name1>": <value>|<list>,
      "<feature_name2>": <value>|<list>,
      ...
    },
    {
      // Example 2
      "<feature_name1>": <value>|<list>,
      "<feature_name2>": <value>|<list>,
      ...
    }
    ...
  ]
}

<value> is a JSON number (whole or decimal), JSON string, or a JSON object that represents binary data (see the Encoding binary values section below for details). <list> is a list of such values. This format is similar to gRPC's ClassificationRequest and RegressionRequest protos. Both versions accept list of Example objects.

Response format

A classify request returns a JSON object in the response body, formatted as follows:

{
  "result": [
    // List of class label/score pairs for first Example (in request)
    [ [<label1>, <score1>], [<label2>, <score2>], ... ],

    // List of class label/score pairs for next Example (in request)
    [ [<label1>, <score1>], [<label2>, <score2>], ... ],
    ...
  ]
}

<label> is a string (which can be an empty string "" if the model does not have a label associated with the score). <score> is a decimal (floating point) number.

The regress request returns a JSON object in the response body, formatted as follows:

{
  // One regression value for each example in the request in the same order.
  "result": [ <value1>, <value2>, <value3>, ...]
}

<value> is a decimal number.

Users of gRPC API will notice the similarity of this format with ClassificationResponse and RegressionResponse protos.

Predict API

This API closely follows the PredictionService.Predict gRPC API.

URL

POST http://host:port/v1/models/${MODEL_NAME}[/versions/${VERSION}|/labels/${LABEL}]:predict

Including /versions/${VERSION} or /labels/${LABEL} is optional. If omitted the latest version is used.

Request format

The request body for predict API must be JSON object formatted as follows:

{
  // (Optional) Serving signature to use.
  // If unspecifed default serving signature is used.
  "signature_name": <string>,

  // Input Tensors in row ("instances") or columnar ("inputs") format.
  // A request can have either of them but NOT both.
  "instances": <value>|<(nested)list>|<list-of-objects>
  "inputs": <value>|<(nested)list>|<object>
}

Specifying input tensors in row format.

This format is similar to PredictRequest proto of gRPC API and the CMLE predict API. Use this format if all named input tensors have the same 0-th dimension. If they don't, use the columnar format described later below.

In the row format, inputs are keyed to instances key in the JSON request.

When there is only one named input, specify the value of instances key to be the value of the input:

{
  // List of 3 scalar tensors.
  "instances": [ "foo", "bar", "baz" ]
}

{
  // List of 2 tensors each of [1, 2] shape
  "instances": [ [[1, 2]], [[3, 4]] ]
}

Tensors are expressed naturally in nested notation since there is no need to manually flatten the list.

For multiple named inputs, each item is expected to be an object containing input name/tensor value pair, one for each named input. As an example, the following is a request with two instances, each with a set of three named input tensors:

{
 "instances": [
   {
     "tag": "foo",
     "signal": [1, 2, 3, 4, 5],
     "sensor": [[1, 2], [3, 4]]
   },
   {
     "tag": "bar",
     "signal": [3, 4, 1, 2, 5]],
     "sensor": [[4, 5], [6, 8]]
   }
 ]
}

Note, each named input ("tag", "signal", "sensor") is implicitly assumed have same 0-th dimension (two in above example, as there are two objects in the instances list). If you have named inputs that have different 0-th dimension, use the columnar format described below.

Specifying input tensors in column format.

Use this format to specify your input tensors, if individual named inputs do not have the same 0-th dimension or you want a more compact representation. This format is similar to the inputs field of the gRPC Predict request.

In the columnar format, inputs are keyed to inputs key in the JSON request.

The value for inputs key can either a single input tensor or a map of input name to tensors (listed in their natural nested form). Each input can have arbitrary shape and need not share the/ same 0-th dimension (aka batch size) as required by the row format described above.

Columnar representation of the previous example is as follows:

{
 "inputs": {
   "tag": ["foo", "bar"],
   "signal": [[1, 2, 3, 4, 5], [3, 4, 1, 2, 5]],
   "sensor": [[[1, 2], [3, 4]], [[4, 5], [6, 8]]]
 }
}

Note, inputs is a JSON object and not a list like instances (used in the row representation). Also, all the named inputs are specified together, as opposed to unrolling them into individual rows done in the row format described previously. This makes the representation compact (but maybe less readable).

Response format

The predict request returns a JSON object in response body.

A request in row format has response formatted as follows:

{
  "predictions": <value>|<(nested)list>|<list-of-objects>
}

If the output of the model contains only one named tensor, we omit the name and predictions key maps to a list of scalar or list values. If the model outputs multiple named tensors, we output a list of objects instead, similar to the request in row-format mentioned above.

A request in columnar format has response formatted as follows:

{
  "outputs": <value>|<(nested)list>|<object>
}

If the output of the model contains only one named tensor, we omit the name and outputs key maps to a list of scalar or list values. If the model outputs multiple named tensors, we output an object instead. Each key of this object corresponds to a named output tensor. The format is similar to the request in column format mentioned above.

Output of binary values

TensorFlow does not distinguish between non-binary and binary strings. All are DT_STRING type. Named tensors that have _bytes as a suffix in their name are considered to have binary values. Such values are encoded differently as described in the encoding binary values section below.

JSON mapping

The RESTful APIs support a canonical encoding in JSON, making it easier to share data between systems. For supported types, the encodings are described on a type-by-type basis in the table below. Types not listed below are implied to be unsupported.

TF Data Type	JSON Value	JSON example	Notes
DT_BOOL	true, false	true, false
DT_STRING	string	"Hello World!"	If `DT_STRING` represents binary bytes (e.g. serialized image bytes or protobuf), encode these in Base64. See Encoding binary values for more info.
DT_INT8, DT_UINT8, DT_INT16, DT_INT32, DT_UINT32, DT_INT64, DT_UINT64	number	1, -10, 0	JSON value will be a decimal number.
DT_FLOAT, DT_DOUBLE	number	1.1, -10.0, 0, `NaN`, `Infinity`	JSON value will be a number or one of the special token values - `NaN`, `Infinity`, and `-Infinity`. See JSON conformance for more info. Exponent notation is also accepted.

Floating Point Precision

JSON has a single number data type. Thus it is possible to provide a value for an input that results in a loss of precision. For instance, if the input x is a float data type, and the input {"x": 1435774380} is sent to the model running on hardware based on the IEEE 754 floating point standard (e.g. Intel or AMD), then the value will be silently converted by the underyling hardware to 1435774336 since 1435774380 cannot be exactly represented in a 32-bit floating point number. Typically, the inputs to serving should be the same distribution as training, so this generally won't be problematic because the same conversions happened at training time. However, in case full precision is needed, be sure to use an underlying data type in your model that can handle the desired precision and/or consider client-side checking.

Encoding binary values

JSON uses UTF-8 encoding. If you have input feature or tensor values that need to be binary (like image bytes), you must Base64 encode the data and encapsulate it in a JSON object having b64 as the key as follows:

{ "b64": <base64 encoded string> }

You can specify this object as a value for an input feature or tensor. The same format is used to encode output response as well.

A classification request with image (binary data) and caption features is shown below:

{
  "signature_name": "classify_objects",
  "examples": [
    {
      "image": { "b64": "aW1hZ2UgYnl0ZXM=" },
      "caption": "seaside"
    },
    {
      "image": { "b64": "YXdlc29tZSBpbWFnZSBieXRlcw==" },
      "caption": "mountains"
    }
  ]
}

JSON conformance

Many feature or tensor values are floating point numbers. Apart from finite values (e.g. 3.14, 1.0 etc.) these can have NaN and non-finite (Infinity and -Infinity) values. Unfortunately the JSON specification (RFC 7159) does NOT recognize these values (though the JavaScript specification does).

The REST API described on this page allows request/response JSON objects to have such values. This implies that requests like the following one are valid:

{
  "example": [
    {
      "sensor_readings": [ 1.0, -3.14, Nan, Infinity ]
    }
  ]
}

A (strict) standards compliant JSON parser will reject this with a parse error (due to NaN and Infinity tokens mixed with actual numbers). To correctly handle requests/responses in your code, use a JSON parser that supports these tokens.

NaN, Infinity, -Infinity tokens are recognized by proto3, Python JSON module and JavaScript language.

Example

We can use the toy half_plus_three model to see REST APIs in action.

Start ModelServer with the REST API endpoint

Download the half_plus_three model from git repository:

$ mkdir -p /tmp/tfserving
$ cd /tmp/tfserving
$ git clone --depth=1 https://github.com/tensorflow/serving

We will use Docker to run the ModelServer. If you want to install ModelServer natively on your system, follow setup instructions to install instead, and start the ModelServer with --rest_api_port option to export REST API endpoint (this is not needed when using Docker).

$ cd /tmp/tfserving
$ docker pull tensorflow/serving:latest
$ docker run --rm -p 8501:8501 \
    --mount type=bind,source=$(pwd),target=$(pwd) \
    -e MODEL_BASE_PATH=$(pwd)/serving/tensorflow_serving/servables/tensorflow/testdata \
    -e MODEL_NAME=saved_model_half_plus_three -t tensorflow/serving:latest
...
.... Exporting HTTP/REST API at:localhost:8501 ...

Make REST API calls to ModelServer

In a different terminal, use the curl tool to make REST API calls.

Get status of the model as follows:

$ curl http://localhost:8501/v1/models/saved_model_half_plus_three
{
 "model_version_status": [
  {
   "version": "123",
   "state": "AVAILABLE",
   "status": {
    "error_code": "OK",
    "error_message": ""
   }
  }
 ]
}

A predict call would look as follows:

$ curl -d '{"instances": [1.0,2.0,5.0]}' -X POST http://localhost:8501/v1/models/saved_model_half_plus_three:predict
{
    "predictions": [3.5, 4.0, 5.5]
}

And a regress call looks as follows:

$ curl -d '{"signature_name": "tensorflow/serving/regress", "examples": [{"x": 1.0}, {"x": 2.0}]}' \
  -X POST http://localhost:8501/v1/models/saved_model_half_plus_three:regress
{
    "results": [3.5, 4.0]
}

Note, regress is available on a non-default signature name and must be specified explicitly. An incorrect request URL or body returns an HTTP error status.

$ curl -i -d '{"instances": [1.0,5.0]}' -X POST http://localhost:8501/v1/models/half:predict
HTTP/1.1 404 Not Found
Content-Type: application/json
Date: Wed, 06 Jun 2018 23:20:12 GMT
Content-Length: 65

{ "error": "Servable not found for request: Latest(half)" }
$