4. Productization

Overview

In this chapter, you will handle new data and create a model that is able to learn continuously from the patterns it is shown and help make better predictions. We will use a web application as an example to show how to deploy deep learning models not only because of the simplicity and prevalence of web apps, but also the different possibilities they provide, such as getting predictions on mobile using a web browser and making REST APIs for users.

Introduction

This chapter focuses on how to productize a deep learning model. We use the word productize to define the creation of a software product from a deep learning model that can be used by other people and applications.

We are interested in models that use new data as and when it becomes available, continuously learn patterns from new data, and consequently, make better predictions. In this chapter, we will study two strategies to deal with new data: one that retrains an existing model, and another that creates a completely new model. Then, we implement the latter strategy in our Bitcoin price prediction model so that it can continuously predict new Bitcoin prices.

By the end of this chapter, we will be able to deploy a working web application (with a functioning HTTP API) and modify it to our heart's content.

Handling New Data

Models can be trained once using a set of data and can then be used to make predictions. Such static models can be very useful, but it is often the case that we want our model to continuously learn from new data—and to continuously get better as it does so.

In this section, we will discuss two strategies of handling new data and see how to implement them in Python.

Separating Data and Model

When building a deep learning application, the two most important areas are data and model. From an architectural point of view, it is recommended that these two areas be kept separate. We believe that is a good suggestion because each of these areas includes functions inherently separate from each other. Data is often required to be collected, cleaned, organized, and normalized, whereas models need to be trained, evaluated, and able to make predictions.

Following that suggestion, we will be using two different code bases to help us build our web application...

Summary

This lesson concludes our journey into creating a deep learning model and deploying it as a web application. Our very last steps included deploying a model that predicts Bitcoin prices built using Keras and the TensorFlow engine. We finished our work by packaging the application as a Docker container and deploying it so that others can consume the predictions of our model, as well as other applications, via its API.

Aside from that work, you have also learned that there is much that can be improved. Our Bitcoin model is only an example of what a model can do (particularly LSTMs). The challenge now is twofold: how can you make that model perform better as time passes? And what features can you add to your web application to make your model more accessible? With the concepts you've learned in this book, you will be able to develop models and keep enhancing them to make accurate predictions.