Generative network understanding

Question

I was going through GAN's notebook by fchallot on Generative Adversarial Networks where, in the Generator Network, he creates a Dense layer with $16*16 * 128$ (where 128 is the number of channels).

How exactly does latent_dim=32  becomes of shape $16 * 16$ in the network. 
How are these values decided?

vico · Answer

latent_dim does not become of shape 16*16

x = layers.Dense(128 * 16 * 16)(generator_input)

mean: The input of size 32 (the latent_dim) is connected to a layer of size 16*16*128

These value are decided by the data scientist, they are the hyper-parameters.
There is a lot of research on how to adjust hyper-parameters for GAN.
People use generally a size of 100 for the latent_dim but it may depends of the complexity of the image you are trying to generate.

Generative network understanding

One Answer

Add your own answers!

Ask a Question