16.03.2021 Views

Advanced Deep Learning with Keras

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Disentangled Representation GANs

The Encoder network is made of a stack of simple encoders, Encoder i

where i = 0 …

n - 1 corresponding to n features. Each encoder extracts certain facial features. For

example, Encoder 0

may be the encoder for hairstyle features, Features 1

. All the simple

encoders contribute to making the overall Encoder perform correct predictions.

The idea behind StackedGAN is that if we would like to build a GAN that generates

fake celebrity faces, we should simply invert the Encoder. StackedGAN are made

of a stack of simpler GANs, GAN i

where i = 0 … n - 1 corresponding to n features.

Each GAN i

learns to invert the process of its corresponding encoder, Encoder i

. For

example, GAN 0

generates fake celebrity faces from fake hairstyle features which is

the inverse of the Encoder 0

process.

Each GAN i

uses a latent code, z i

, that conditions its generator output. For example,

the latent code, z 0

, can alter the hairstyle from curly to wavy. The stack of GANs

can also act as one to synthesize fake celebrity faces, completing the inverse process

of the whole Encoder. The latent code of each GAN i

, z i

, can be used to alter specific

attributes of fake celebrity faces:

Figure 6.2.1: The basic idea of StackedGAN in the context of celebrity faces generation. Assuming that there is

a hypothetical deep encoder network that can perform classification on celebrity faces, a StackedGAN simply

inverts the process of the encoder.

[ 180 ]

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!