Serving a TensorFlow Model

This tutorial shows you how to use TensorFlow Serving components to export a trained TensorFlow model and use the standard tensorflow_model_server to serve it. If you are already familiar with TensorFlow Serving, and you want to know more about how the server internals work, see the TensorFlow Serving advanced tutorial.

This tutorial uses a simple Softmax Regression model that classifies handwritten digits. It is very similar to the one introduced in the TensorFlow tutorial on image classification using the Fashion MNIST dataset.

The code for this tutorial consists of two parts:

  • A Python file,, that trains and exports the model.

  • A ModelServer binary which can be either installed using Apt, or compiled from a C++ file ( The TensorFlow Serving ModelServer discovers new exported models and runs a gRPC service for serving them.

Before getting started, first install Docker.

Train and export TensorFlow model

For the training phase, the TensorFlow graph is launched in TensorFlow session sess, with the input tensor (image) as x and output tensor (Softmax score) as y.

Then we use TensorFlow's SavedModelBuilder module to export the model. SavedModelBuilder saves a "snapshot" of the trained model to reliable storage so that it can be loaded later for inference.

For details on the SavedModel format, please see the documentation at SavedModel

From, the following is a short code snippet to illustrate the general process of saving a model to disk.

export_path_base = sys.argv[-1]
export_path = os.path.join(
print('Exporting trained model to', export_path)
builder = tf.saved_model.builder.SavedModelBuilder(export_path)
    sess, [tf.compat.v1.saved_model.tag_constants.SERVING],

SavedModelBuilder.__init__ takes the following argument:

  • export_path is the path of the export directory.

SavedModelBuilder will create the directory if it does not exist. In the example, we concatenate the command line argument and FLAGS.model_version to obtain the export directory. FLAGS.model_version specifies the version of the model. You should specify a larger integer value when exporting a newer version of the same model. Each version will be exported to a different sub-directory under the given path.

You can add meta graph and variables to the builder using SavedModelBuilder.add_meta_graph_and_variables() with the following arguments:

  • sess is the TensorFlow session that holds the trained model you are exporting.

  • tags is the set of tags with which to save the meta graph. In this case, since we intend to use the graph in serving, we use the serve tag from predefined SavedModel tag constants. For more details, see and related TensorFlow API documentation.

  • signature_def_map specifies the map of user-supplied key for a signature to a tensorflow::SignatureDef to add to the meta graph. Signature specifies what type of model is being exported, and the input/output tensors to bind to when running inference.

    The special signature key serving_default specifies the default serving signature. The default serving signature def key, along with other constants related to signatures, are defined as part of SavedModel signature constants. For more details, see and related TensorFlow API documentation.

    Further, to help build signature defs easily, the SavedModel API provides signature def utils.. Specifically, in the original file, we use signature_def_utils.build_signature_def() to build predict_signature and classification_signature.

    As an example for how predict_signature is defined, the util takes the following arguments:

    • inputs={'images': tensor_info_x} specifies the input tensor info.

    • outputs={'scores': tensor_info_y} specifies the scores tensor info.

    • method_name is the method used for the inference. For Prediction requests, it should be set to tensorflow/serving/predict. For other method names, see and related TensorFlow API documentation.

Note that tensor_info_x and tensor_info_y have the structure of tensorflow::TensorInfo protocol buffer defined here. To easily build tensor infos, the TensorFlow SavedModel API also provides, with related TensorFlow API documentation.

Also, note that images and scores are tensor alias names. They can be whatever unique strings you want, and they will become the logical names of tensor x and y that you refer to for tensor binding when sending prediction requests later.

For instance, if x refers to the tensor with name 'long_tensor_name_foo' and y refers to the tensor with name 'generated_tensor_name_bar', builder will store tensor logical name to real name mapping ('images' -> 'long_tensor_name_foo') and ('scores' -> 'generated_tensor_name_bar'). This allows the user to refer to these tensors with their logical names when running inference.

Let's run it!

First, if you haven't done so yet, clone this repository to your local machine:

git clone
cd serving

Clear the export directory if it already exists:

rm -rf /tmp/mnist

Now let's train the model:

tools/ python tensorflow_serving/example/ \

This should result in output that looks like:

Training model...


Done training!
Exporting trained model to models/mnist
Done exporting!

Now let's take a look at the export directory.

$ ls /tmp/mnist

As mentioned above, a sub-directory will be created for exporting each version of the model. FLAGS.model_version has the default value of 1, therefore the corresponding sub-directory 1 is created.

$ ls /tmp/mnist/1
saved_model.pb variables

Each version sub-directory contains the following files:

  • saved_model.pb is the serialized tensorflow::SavedModel. It includes one or more graph definitions of the model, as well as metadata of the model such as signatures.

  • variables are files that hold the serialized variables of the graphs.

With that, your TensorFlow model is exported and ready to be loaded!

Load exported model with standard TensorFlow ModelServer

Use a Docker serving image to easily load the model for serving:

docker run -p 8500:8500 \
--mount type=bind,source=/tmp/mnist,target=/models/mnist \
-e MODEL_NAME=mnist -t tensorflow/serving &

Test the server

We can use the provided mnist_client utility to test the server. The client downloads MNIST test data, sends them as requests to the server, and calculates the inference error rate.

tools/ python tensorflow_serving/example/ \
  --num_tests=1000 --server=

This should output something like

    Inference error rate: 11.13%

We expect around 90% accuracy for the trained Softmax model and we get 11% inference error rate for the first 1000 test images. This confirms that the server loads and runs the trained model successfully!