Understanding DiscoGANs

In this section, we are mainly going to take a closer look at Discovery GANS, which are popularly known as DiscoGANs.

Before going further, let's try to understand reconstruction loss in machine learning, since this is one of the concepts that this chapter is majorly dependent on. When learning about the representation of an unstructured data type such as an image/text, we want our model to encode the data in such a manner that when it's decoded, the underlying image/text can be generated back. To incorporate this condition in the model explicitly, we use a reconstruction loss (essentially the Euclidean distance between the reconstructed and original image) in training the model.

Style transfer has been one of the most prominent use cases of GANs. Style transfer basically refers to the problem where, if you are given an image/data in one domain, is it possible to successfully generate an image/data in another domain. This problem has become quite famous among several...