Transfer learning in Keras

Unlike in other examples in this book, here we will need to cover both the target domain problem, the source domain problem, and the network architecture we're using. We will start with an overview of the target domain, which is the problem we're trying to solve. Then we will cover the source domain our network was originally trained on and briefly cover the network architecture we will be using. Then, we will spend the rest of the chapter wiring the problem together. We need to consider both domains separately because their size and similarity are closely related to network performance. The closer the target and source are in type, the better the results.

Target domain overview

In this chapter's example, I will be working with Kaggle's Dogs versus Cats dataset. This dataset consists of 25,000 images of dogs and cats. It's perfectly balanced between classes at 12,500 each. The dataset can be downloaded from https://www.kaggle.com/c/dogs-vs-cats/data.

This is a binary...