Learning TensorFlow

Chapter 10. Exporting and Serving Models with TensorFlow

In this chapter we will learn how to save and export models by using both simple and advanced production-ready methods. For the latter we introduce TensorFlow Serving, one of TensorFlow’s most practical tools for creating production environments. We start this chapter with a quick overview of two simple ways to save models and variables: first by manually saving the weights and reassigning them, and then by using the Saver class that creates training checkpoints for our variables and also exports our model. Finally, we shift to more advanced applications where we can deploy our model on a server by using TensorFlow Serving.

Saving and Exporting Our Model

So far we’ve dealt with how to create, train, and track models with TensorFlow. Now we will see how to save a trained model. Saving the current state of our weights is crucial for obvious practical reasons—we don’t want to have to retrain our model from scratch every time, and we also want a convenient way to share the state of our model with others (as in the pretrained models we saw in Chapter 7).

In this section we go over the basics of saving and exporting. We start with a simple way of saving and loading our weights to and from files. Then we will see how to use TensorFlow’s Saver object to keep serialized model checkpoints that include information about both the state of our weights and our constructed graph.     

Assigning Loaded Weights

A naive but practical way ...