Chapter 3. Cats Versus Dogs: Transfer Learning in 30 Lines with Keras
Imagine that we want to learn how to play the melodica, a wind instrument in the form of a handheld keyboard. Without a musical background, and the melodica being our very first instrument, it might take us a few months to become proficient at playing it. In contrast, if we were already skilled at playing another instrument, such as the piano, it might just take a few days, given how similar the two instruments are. Taking the learnings from one task and fine tuning them on a similar task is something we often do in real life (as illustrated in Figure 3-1). The more similar the two tasks are, the easier it is to adapt the learning from one task to the other.
We can apply this phenomenon from real life to the world of deep learning. Starting a deep learning project can be relatively quick when using a pretrained model, which reuses the knowledge that it learned during its training, and adapt it to the task at hand. This process is known as transfer learning.
In this chapter, we use transfer learning to modify existing models by training our own classifier in minutes using Keras. By the end of this chapter, we will have several tools in our arsenal to create high-accuracy image classifiers for any task.
Adapting Pretrained Models to New Tasks
Before we discuss the process ...