Implementing a U-Net with transfer learning

Training a U-Net from scratch is a very good first step toward creating a performant image segmentation system. However, one of the biggest superpowers in deep learning that's applied to computer vision is being able to build solutions on top of the knowledge of other networks, which usually leads to faster and better results.

Image segmentation is no exception to this rule, and in this recipe, we'll implement a better segmentation network using transfer learning.

Let's begin.

Getting ready

This recipe is very similar to the previous one (Implementing a U-Net from scratch), so we'll only go into depth on the parts that are different. For a deeper explanation, I recommend that you complete the Implementing a U-Net from scratch recipe before attempting this one. As expected, the libraries we'll need are the same as they were for that recipe, all of which can be installed using pip. Let's start with...