Assembling the pieces to deploy machine learning systems
Getting machine learning systems to really work requires excellent machine learning models and algorithms, but also often requires lots of good systems work to surround the core machine learning models/algorithms to make them useful as components in larger systems. We've been building up these pieces and open-sourcing them to make it easier for everyone to use machine learning in their products and applications.
A while ago, I posted (post: https://plus.google.com/+JeffDean/posts/6okdnD1MHmX
) about the open-sourcing of TensorFlow Serving, an open source package developed at Google that complements the core TensorFlow system. TensorFlow Serving makes it easy to take models that have been trained with TensorFlow and move them into a system for serving inference requests on those models (and simplifying messy issues like updating models in the live serving system as they are updated through continuous training, etc.).
This blog post by the TensorFlow Serving team at Google shows how to use TensorFlow, TensorFlow Serving, and Kubernetes (another project open-sourced by Google) to deploy a pre-trained Inception-v3 image classification model using TensorFlow Serving, running in Kubernetes containers, and using Kubernetes' ability to dynamically scale the number of replicas up and down as load changes.
Full tutorial link from the bottom of the blog post: https://tensorflow.github.io/serving/serving_inception
Class(ification) is starting!