As far as I've understood, the architecture of the generator in this implementation mainly consists of two fully-connected layers and two "deconvolutional" ones.
However, in the DCGAN paper, the generator consists of four deconvolutional layers (more correctly would be saying "fractionally-strided convolutions").
Why such choice?
As far as I've understood, the architecture of the generator in this implementation mainly consists of two fully-connected layers and two "deconvolutional" ones. However, in the DCGAN paper, the generator consists of four deconvolutional layers (more correctly would be saying "fractionally-strided convolutions"). Why such choice?