Skip to content
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
183 lines (158 sloc) 10.4 KB

Image Recognition with TensorFlow Serving on CPU

Real-time and Max Throughput Inference

Models: ResNet50, InceptionV3


This tutorial will introduce you to the CPU performance considerations for image recognition deep learning models and how to use Intel® Optimizations for TensorFlow Serving to improve inference time on CPUs. It also provides sample code that you can use to get your optimized TensorFlow model server and GRPC client up and running quickly.


This tutorial assumes you have already:


Convolutional neural networks (CNNs) for image recognition are computationally expensive. The Intel® Math Kernel Library for Deep Neural Networks (Intel® MKL-DNN) offers significant performance improvements for convolution, pooling, normalization, activation, and other operations via efficient vectorization and multi-threading. Tuning TensorFlow Serving to take full advantage of your hardware for image recognition deep learning inference involves:

  1. Working through this tutorial to set up servable versions of the well-known ResNet50 and InceptionV3 CNN models
  2. Running a TensorFlow Serving docker container configured for performance given your hardware resources
  3. Running a client script to measure latency and throughput
  4. Experimenting with the TensorFlow Serving settings on your own to further optimize for your model and use case

Hands-on Tutorial - ResNet50 or InceptionV3

For steps 1 and 2, refer to the Intel Model Zoo FP32 benchmarks:

  1. Download the Model: Download and extract the ResNet50 or InceptionV3 pre-trained model (FP32), using the instructions in one of the READMEs above.

  2. (Optional) Download Data: If you are interested only in testing latency and throughput, not accuracy, you can skip this step and use synthetic data. If you want to verify prediction accuracy by testing on real data, follow the instructions in one of the READMEs above to download the ImageNet dataset.

  3. Clone this repository: Clone the intelai/models repository and cd into the docs/image_recognition/tensorflow_serving/src directory.

    $ git clone
    $ cd models/docs/image_recognition/tensorflow_serving/src
  4. Set up your environment: In this tutorial, we use a virtual environment to install a few required Python packages. If you do not have pip or virtualenv, you will need to get them first:

    $ sudo apt-get install python-pip
    $ pip install virtualenv
    $ virtualenv venv

    Then activate the virtual environment and install grpc, requests, tensorflow, and tensorflow-serving-api (at the time of this writing, the order of installation matters):

    $ source venv/bin/activate
    (venv)$ pip install grpc
    (venv)$ pip install requests
    (venv)$ pip install intel-tensorflow
    (venv)$ pip install tensorflow-serving-api
  5. Create a SavedModel: Using the conversion script, convert the pre-trained model graph to a SavedModel.


    (venv)$ python --import_path inceptionv3_fp32_pretrained_model.pb
    2018-12-11 15:55:33.018355: I tensorflow/core/platform/] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 AVX512F FMA
    2018-12-11 15:55:33.033707: I tensorflow/core/common_runtime/] Creating new thread pool with default inter op setting: 2. Tune using inter_op_parallelism_threads for best performance.
    2018-12-11 15:55:33.447678: I tensorflow/tools/graph_transforms/] Applying strip_unused_nodes
    2018-12-11 15:55:33.515980: I tensorflow/tools/graph_transforms/] Applying remove_nodes
    2018-12-11 15:55:33.556799: I tensorflow/tools/graph_transforms/] Applying fold_constants
    2018-12-11 15:55:33.575557: I tensorflow/tools/graph_transforms/] Applying fold_batch_norms
    2018-12-11 15:55:33.586407: I tensorflow/tools/graph_transforms/] Applying fold_old_batch_norms
    Exporting trained model to /tmp/1

    This will create a /tmp/1/ directory with a saved_model.pb file in it. This is the file we will serve from TensorFlow Serving. The script has applied some transform optimizations and attached a signature definition to the model in order to make it compatible with TensorFlow Serving. You can take a look at the script, its flags/options, and these resources for more information:

  6. Discover the number of physical cores: Compute num_physical_cores by executing the lscpu command and multiplying Core(s) per socket by Socket(s). For example, for a machine with Core(s) per socket: 28 and Socket(s): 2, num_physical_cores = 28 * 2 = 56.

  7. Start the server: Now let's start up the TensorFlow model server. To optimize overall performance, use the following recommended settings from the General Best Practices:

    • OMP_NUM_THREADS=num_physical_cores
    • TENSORFLOW_INTRA_OP_PARALLELISM=num_physical_cores

    For our example with 56 physical cores:

    (venv)$ docker run \
            --name=tfserving \
            --rm \
            -d \
            -p 8500:8500 \
            -v "/tmp:/models/inceptionv3" \
            -e MODEL_NAME=inceptionv3 \
            -e OMP_NUM_THREADS=56 \

    Note: For some models, playing around with these settings values can improve performance even further. We recommend that you experiment with your own hardware and model if you have strict performance requirements.

  8. Run a Test: Now we can run a test client that downloads a cat picture and sends it for recognition. The script has an option for sending a local JPG, if you would prefer to test a different image. Run python --help for more usage information.

    (venv)$ python --model inceptionv3

    The output should be a tensor of class probabilities and Predicted class: 286.

  9. Real-time inference: Real-time inference is measured by latency and is usually defined as batch size 1. To see average inference latency (in ms), run the benchmark script using batch_size 1:

    (venv)$ python --batch_size 1 --model inceptionv3
    Iteration 1: 0.017 sec
    Iteration 40: 0.016 sec
    Average time: 0.016 sec
    Batch size = 1
    Latency: 16.496 ms
    Throughput: 60.619 images/sec

    In some cases, it is desirable to constrain the inference server to a single core or socket. Docker has many runtime flags that allow you to control the container's access to the host system's CPUs, memory, and other resources. See the Docker document on this topic for all the options and their definitions. For example, to run the container so that a single CPU is used, you can use these settings:

    • --cpuset-cpus="0"
    • --cpus="1"
    (venv)$ docker run \
            --name=tfserving \
            --rm \
            --cpuset-cpus="0" \
            --cpus="1" \
            -d \
            -p 8500:8500 \
            -v "/tmp:/models/inceptionv3" \
            -e MODEL_NAME=inceptionv3 \
            -e OMP_NUM_THREADS=1 \
  10. Maximum throughput: Regardless of hardware, the best batch size for throughput is 128. To see average throughput (in images/sec), run the benchmark script using batch_size 128:

    (venv)$ python --batch_size 128 --model inceptionv3
    Iteration 1: 1.706 sec
    Iteration 40: 0.707 sec
    Average time: 0.693 sec
    Batch size = 128
    Throughput: 184.669 images/sec
  11. Clean up:

    • To shut off the server, stop the docker process that is running it. View your running containers with docker ps. To stop one, copy the Container ID and run docker stop <container_id>.
    • Deactivate your virtual environment with deactivate.


You have now seen two end-to-end examples of serving an image recognition model for inference using TensorFlow Serving, and learned:

  1. How to create a SavedModel from a TensorFlow model graph
  2. How to choose good values for the performance-related runtime parameters exposed by the docker run command
  3. How to verify that the served model can correctly classify an image using a GRPC client
  4. How to benchmark latency and throughput metrics using a GRPC client

With this knowledge and the example code provided, you should be able to get started serving your own custom image recognition model with good performance. If desired, you should also be able to investigate a variety of different settings combinations to see if further performance improvement are possible.

You can’t perform that action at this time.