Skip to content
This repository has been archived by the owner on Jul 22, 2024. It is now read-only.

Latest commit

 

History

History
69 lines (57 loc) · 4.63 KB

README.md

File metadata and controls

69 lines (57 loc) · 4.63 KB
Raw

Demonstrating interaction between z/OS and TensorFlow Serving

This project intends to demonstrate a call to TensorFlow Serving using REST API from a simple z/OS Java app. The purpose is to create a simple and easily deployable scenario that can be used on z/OS to understand serving concepts.

This model performns a matrix multiplication of a [1,5] input tensor by a [5,1] weights tensor, where all weights are defined with a value of 1. This results in an output tensor shape of [1,1].

For example, with this model an input tensor of [[1,2,3,4,5]] is multiplied by a weights tensor of [[1],[1],[1],[1],[1]], resulting in a value of [[15.]]

Note that since this project is intended for deployment to z/OS, we are avoiding managing dependecies through sub-moduling or Maven/Ant. The intent is to quickly try this project out without installing additional software. Required Jar files can be pulled from the references below.

Deploying TensorFlow Serving on z/OS Container Extensions

  • TensorFlow Serving can be obtained from the IBM Container Image Repository or built from source for Linux on Z

  • Details on the TFServing API can be found here: https://www.tensorflow.org/tfx/serving/api_rest

    • This example assumes a TensorFlow Serving V2.x REST API and has been tested up to TensorFlow 2.7

  • To deploy the example model, you can follow this procedure:

    • SFTP the TensorFlow saved model to zCX. This should be the model folder and all subfolders.
      • The saved model can be found in the model subdirectory of this project.
      • The notebook used to create the model is simplemm.ipynb
    • Create a new docker volume.
      • e.g., docker volume create tfmodels
      • tfmodels is the volume name we create to use in subsequent steps
    • Copy the model directory into the docker volume.
      • One approach is to create a simple container using the volume to allow a docker cp:
        • docker container create --name tfsimple -v tfmodels:/models simple_image
        • docker cp <host_model_dir> tfsimple:/models
          • tfsimple is the container name we create to facilitate the copy via docker CP.
          • simple_image can be any base image, and it can be deleted after this copy.
          • models is a directory we choose to copy the model into.
    • Run the TFServing image: - docker run -d --rm -p 8507:8501 -v tfmodels:/models -e MODEL_NAME=<model_name> <image id> - 8501 is the default TFServing REST port. Here it is mapped to zCX port 8507. - model_name is the TensorFlow model name; this is commonly the directory name that holds the saved_model.pb file

Updating and deploying the z/OS application

  • First, deploy the project to your host system. This program is intended for Unix environments.

  • The following Jar files must be in the project root directory;

  • run make to build compile the .java files

  • issue the following command to try a TensorFlow Serving request:

    • make run HOSTIP=tf-serve-ip PORT=tf-serv-port MODEL_DIR=model-dir PAYLOAD=input-tensor
      • tf-serve-ip is the IP address of the server or instance hosting TensorFlow Serving
      • tf-serv-port is the TF Serving REST port
      • model-dir is the model path, which should typically be /v1/models/simplemm for the sample model.
      • input-tensor consists of 5 comma separated decimal values.

    For example, if we mapped the TFServing REST port to 8507, we would use something like this (changing the IP address to one in use by zCX): make run HOSTIP=127.0.0.1 PORT=8507 MODEL_DIR=/v1/models/simplemm PAYLOAD=1.0,2.0,3.0,4.0,5.0

    In addition to verbose messages, the result for test would show: { "predictions": [[15.0] ] }