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.

Variational Autoencoders (VAEs)

Figures 8.3.1 and 8.3.2 show the latent vector means for β -VAE with β = 7 and β = 10 .

With β = 7 , the distribution has a smaller standard deviation when compared to

CVAE. With β = 10 , there is only the latent code that is learned. The distribution

is practically shrunk to 1D with the first latent code z[0] ignored by the encoder

and decoder:

Figure 8.3.2: The latent vector mean values for the test dataset ( β -VAE with )

Color images can be found on the book GitHub repository: https://github.com/PacktPublishing/Advanced-

Deep-Learning-with-Keras/tree/master/chapter8-vae.

These observations are reflected in Figure 8.3.3. β -VAE with β = 7 has two latent

codes that are practically independent. z[0] determines the tilt of the writing style.

Meanwhile, z[1] specifies the width and roundness (if applicable) of the digits.

For β -VAE with β = 10 , z[0] is muted. Increasing z[0] does not alter the digit in

a significant way. z[1] determines the tilt angle and width of the writing style.

[ 266 ]

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

Saved successfully!

Ooh no, something went wrong!