Using TensorFlow Serving via Docker

One of the easiest ways to get started using TensorFlow Serving is via Docker.

Installing Docker

General installation instructions are on the Docker site, but we give some quick links here:

Serving with Docker

Pulling a serving image

Once you have Docker installed, you can pull the latest TensorFlow Serving docker image by running:

docker pull tensorflow/serving

This will pull down an minimal Docker image with TensorFlow Serving installed.

See the Docker Hub tensorflow/serving repo for other versions of images you can pull.

Running a serving image

The serving images (both CPU and GPU) have the following properties:

  • Port 8500 exposed for gRPC
  • Port 8501 exposed for the REST API
  • Optional environment variable MODEL_NAME (defaults to model)
  • Optional environment variable MODEL_BASE_PATH (defaults to /models)

When the serving image runs ModelServer, it runs it as follows:

tensorflow_model_server --port=8500 --rest_api_port=8501 \
  --model_name=${MODEL_NAME} --model_base_path=${MODEL_BASE_PATH}/${MODEL_NAME}

To serve with Docker, you'll need:

  • An open port on your host to serve on
  • A SavedModel to serve
  • A name for your model that your client will refer to

What you'll do is run the Docker container, publish the container's ports to your host's ports, and mounting your host's path to the SavedModel to where the container expects models.

Let's look at an example:

docker run -p 8501:8501 \
  --mount type=bind,source=/path/to/my_model/,target=/models/my_model \
  -e MODEL_NAME=my_model -t tensorflow/serving

In this case, we've started a Docker container, published the REST API port 8501 to our host's port 8501, and taken a model we named my_model and bound it to the default model base path (${MODEL_BASE_PATH}/${MODEL_NAME} = /models/my_model). Finally, we've filled in the environment variable MODEL_NAME with my_model, and left MODEL_BASE_PATH to its default value.

This will run in the container:

tensorflow_model_server --port=8500 --rest_api_port=8501 \
  --model_name=my_model --model_base_path=/models/my_model

If we wanted to publish the gRPC port, we would use -p 8500:8500. You can have both gRPC and REST API ports open at the same time, or choose to only open one or the other.

Passing additional arguments

tensorflow_model_server supports many additional arguments that you could pass to the serving docker containers. For example, if we wanted to pass a model config file instead of specifying the model name, we could do the following:

docker run -p 8500:8500 8501:8501 \
  --mount type=bind,source=/path/to/my_model/,target=/models/my_model \
  --mount type=bind,source=/path/to/my/models.config,target=/models/models.config \
  -t tensorflow/serving --model_config_file=/models/models.config

This approach works for any of the other command line arguments that tensorflow_model_server supports.

Creating your own serving image

If you want a serving image that has your model built into the container, you can create your own image.

First run a serving image as a daemon:

docker run -d --name serving_base tensorflow/serving

Next, copy your SavedModel to the container's model folder:

docker cp models/<my model> serving_base:/models/<my model>

Finally, commit the container that's serving your model by changing MODEL_NAME to match your model's name `':

docker commit --change "ENV MODEL_NAME <my model>" serving_base <my container>

You can now stop serving_base

docker kill serving_base

This will leave you with a Docker image called <my container> that you can deploy and will load your model for serving on startup.

Serving example

Let's run through a full example where we load a SavedModel and call it via the REST API. First pull the serving image:

docker pull tensorflow/serving

This will pull the latest TensorFlow Serving image with ModelServer installed.

Next, we will use a toy model called Half Plus Two, which generates 0.5 * x + 2 for the values of x we provide for prediction.

To get this model, first clone the TensorFlow Serving repo.

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

Next, run the TensorFlow Serving container pointing it to this model and opening the REST API port (8501):

docker run -p 8501:8501 \
  --mount type=bind,\
  source=/tmp/tfserving/serving/tensorflow_serving/servables/tensorflow/testdata/saved_model_half_plus_two_cpu,\
  target=/models/half_plus_two \
  -e MODEL_NAME=half_plus_two -t tensorflow/serving &

This will run the docker container and launch the TensorFlow Serving Model Server, bind the REST API port 8501, and map our desired model from our host to where models are expected in the container. We also pass the name of the model as an environment variable, which will be important when we query the model.

To query the model using the predict API, you can run

curl -d '{"instances": [1.0, 2.0, 5.0]}' \
  -X POST http://localhost:8501/v1/models/half_plus_two:predict

NOTE: Older versions of Windows and other systems without curl can download it here.

This should return a set of values:

{ "predictions": [2.5, 3.0, 4.5] }

More information on using the RESTful API can be found here.

Serving with Docker using your GPU

Install nvidia-docker

Before serving with a GPU, in addition to installing Docker, you will need:

Running a GPU serving image

Running a GPU serving image is identical to running a CPU image. For more details, see running a serving image.

GPU Serving example

Let's run through a full example where we load a model with GPU-bound ops and call it via the REST API.

First install nvidia-docker. Next you can pull the latest TensorFlow Serving GPU docker image by running:

docker pull tensorflow/serving:latest-gpu

This will pull down an minimal Docker image with ModelServer built for running on GPUs installed.

Next, we will use a toy model called Half Plus Two, which generates 0.5 * x + 2 for the values of x we provide for prediction. This model will have ops bound to the GPU device, and will not run on the CPU.

To get this model, first clone the TensorFlow Serving repo.

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

Next, run the TensorFlow Serving container pointing it to this model and opening the REST API port (8501):

docker run --runtime=nvidia -p 8501:8501 \
  --mount type=bind,\
  source=/tmp/tfserving/serving/tensorflow_serving/servables/tensorflow/testdata/saved_model_half_plus_two_gpu,\
  target=/models/half_plus_two \
  -e MODEL_NAME=half_plus_two -t tensorflow/serving:latest-gpu &

This will run the docker container with the nvidia-docker runtime, launch the TensorFlow Serving Model Server, bind the REST API port 8501, and map our desired model from our host to where models are expected in the container. We also pass the name of the model as an environment variable, which will be important when we query the model.

TIP: Before querying the model, be sure to wait till you see a message like the following, indicating that the server is ready to receive requests:

2018-07-27 00:07:20.773693: I tensorflow_serving/model_servers/main.cc:333]
Exporting HTTP/REST API at:localhost:8501 ...

To query the model using the predict API, you can run

$ curl -d '{"instances": [1.0, 2.0, 5.0]}' \
  -X POST http://localhost:8501/v1/models/half_plus_two:predict

NOTE: Older versions of Windows and other systems without curl can download it here.

This should return a set of values:

{ "predictions": [2.5, 3.0, 4.5] }

TIP: Trying to run the GPU model on a machine without a GPU or without a working GPU build of TensorFlow Model Server will result in an error that looks like:

Cannot assign a device for operation 'a': Operation was explicitly assigned to /device:GPU:0

More information on using the RESTful API can be found here.

Developing with Docker

Pulling a development image

For a development environment where you can build TensorFlow Serving, you can try:

docker pull tensorflow/serving:latest-devel

For a development environment where you can build TensorFlow Serving with GPU support, use:

docker pull tensorflow/serving:latest-devel-gpu

See the Docker Hub tensorflow/serving repo for other versions of images you can pull.

Development example

After pulling one of the development Docker images, you can run it while opening the gRPC port (8500):

docker run -it -p 8500:8500 tensorflow/serving:latest-devel

Testing the development environment

To test a model, from inside the container try:

# build the exporter
bazel build -c opt //tensorflow_serving/example:mnist_saved_model
# train the mnist model
bazel-bin/tensorflow_serving/example/mnist_saved_model /tmp/mnist_model
# serve the model
tensorflow_model_server --port=8500 --model_name=mnist --model_base_path=/tmp/mnist_model/ &
# build the client
bazel build -c opt //tensorflow_serving/example:mnist_client
# test the client
bazel-bin/tensorflow_serving/example/mnist_client --num_tests=1000 --server=localhost:8500

Dockerfiles

We currently maintain the following Dockerfiles:

  • Dockerfile, which is a minimal VM with TensorFlow Serving installed.

  • Dockerfile.gpu, which is a minimal VM with TensorFlow Serving with GPU support to be used with nvidia-docker.

  • Dockerfile.devel, which is a minimal VM with all of the dependencies needed to build TensorFlow Serving.

  • Dockerfile.devel-gpu, which is a minimal VM with all of the dependencies needed to build TensorFlow Serving with GPU support.

Building a container from a Dockerfile

If you'd like to build your own Docker image from a Dockerfile, you can do so by running the Docker build command:

Dockerfile:

docker build --pull -t $USER/tensorflow-serving .

Dockerfile.gpu:

docker build --pull -t $USER/tensorflow-serving-gpu -f Dockerfile.gpu .

Dockerfile.devel:

docker build --pull -t $USER/tensorflow-serving-devel -f Dockerfile.devel .

Dockerfile.devel-gpu:

docker build --pull -t $USER/tensorflow-serving-devel-gpu -f Dockerfile.devel-gpu .

TIP: Before attempting to build an image, check the Docker Hub tensorflow/serving repo to make sure an image that meets your needs doesn't already exist.

Building from sources consumes a lot of RAM. If RAM is an issue on your system, you may limit RAM usage by specifying --local_resources=2048,.5,1.0 while invoking Bazel. See the Bazel docs for more information. You can use this same mechanism to tweak the optmizations you're building TensorFlow Serving with. For example:

docker build --pull --build-arg TF_SERVING_BUILD_OPTIONS="--copt=-mavx \
  --cxxopt=-D_GLIBCXX_USE_CXX11_ABI=0 --local_resources 2048,.5,1.0" -t \
  $USER/tensorflow-serving-devel -f Dockerfile.devel .

Running a container

This assumes you have built the Dockerfile.devel container.

To run the container opening the gRPC port (8500):

docker run -it -p 8500:8500 $USER/tensorflow-serving-devel

TIP: If you're running a GPU image, be sure to run using the NVIDIA runtime --runtime=nvidia.

From here, you can follow the instructions for testing a development environment.

Building an optimized serving binary

When running TensorFlow Serving's ModelServer, you may notice a log message that looks like this:

I external/org_tensorflow/tensorflow/core/platform/cpu_feature_guard.cc:141]
Your CPU supports instructions that this TensorFlow binary was not compiled to
use: AVX2 FMA

This indicates that your ModelServer binary isn't fully optimized for the CPU its running on. Depending on the model you are serving, further optimizations may not be necessary. However, building an optimized binary is straight-forward.

When building a Docker image from the provided Dockerfile.devel or Dockerfile.devel-gpu files, the ModelServer binary will be built with the flag -march=native. This will cause Bazel to build a ModelServer binary with all of the CPU optimizations the host you're building the Docker image on supports.

To create a serving image that's fully optimized for your host, simply:

  1. Clone the TensorFlow Serving project

    git clone https://github.com/tensorflow/serving
    cd serving
    
  2. Build an image with an optimized ModelServer

    • For CPU:

      docker build --pull -t $USER/tensorflow-serving-devel \
        -f tensorflow_serving/tools/docker/Dockerfile.devel .
      
    • For GPU: `

      docker build --pull -t $USER/tensorflow-serving-devel-gpu \
        -f tensorflow_serving/tools/docker/Dockerfile.devel-gpu .
      
  3. Build a serving image with the development image as a base

    • For CPU:

      docker build -t $USER/tensorflow-serving \
        --build-arg TF_SERVING_BUILD_IMAGE=$USER/tensorflow-serving-devel \
        -f tensorflow_serving/tools/docker/Dockerfile .
      

      Your new optimized Docker image is now $USER/tensorflow-serving, which you can use just as you would the standard tensorflow/serving:latest image.

    • For GPU:

      docker build -t $USER/tensorflow-serving-gpu \
        --build-arg TF_SERVING_BUILD_IMAGE=$USER/tensorflow-serving-devel-gpu \
        -f tensorflow_serving/tools/docker/Dockerfile.gpu .
      

      Your new optimized Docker image is now $USER/tensorflow-serving-gpu, which you can use just as you would the standard tensorflow/serving:latest-gpu image.