Awesome Chopstick Classifier. TensorFlow Serving examples for Kubernetes

This project proveides examples on different ways to train, export and serve TensorFlow models using Docker, Kubernetes, and TensorFlow Serving. This repository contains conpanion code for the tutorial LINK TBD.

Prerequisites

• Docker
• Kubernetes. Consider using Minikube for local experiments and kubeadm for simple cluster set up.
• For model training and export: Python 3, pip install -r ./py3-train-requirements.txt
• For running the client: Python 2, pip install -r ./py2-client-requirements.txt

Building and running the TensorFlow Serving image

You may use provided Makefile to

• Build the image with make tfserve_image. Notice that no compilation isperformed through the process, the image uses precompiled TensorFlow Serving Ubuntu package.
• Train the classifier and export it with make train_classifier
• Run TensorFlow Serving via docker with make run_server

Using different APIs

TensorFlow Serving examples are provided both for TensorFlow Estimator API and for plain TensorFlow clode since approaches for model export differ significantly between those.

You can use API_TO_USE variable to set the API of interest (the default one is estimator_api):

make API_TO_USE=tf_api clean train_classifier
make API_TO_USE=estimator_api clean train_classifier

Calling example client

You must use python 2 on the client side as the tensorflow-serving-api package is only provided for Python 2 at the current moment. You may call the example client using this command:

python2 client.py --model-name tf_model localhost:8500 0.5

Execute python2 client.py -h if you want to kow more about the parameters.