As mentioned previously, normal GANs have a generator and a discriminator. Let's try to understand the building blocks of DiscoGANs and then proceed to understand how to combine them so that we can learn about cross-domain relationships. These are as follows:
- Generator: In the original GANs, the generator would take an input vector z randomly sampled from, say, Gaussian distribution, and generate fake images. In this case, however, since we are looking to transfer images from one domain to another, we replace the input vector z with an image. Here are the parameters of the generator function:
|
Parameters |
Value |
|
Input image size |
64x64x3 |
|
Output image size |
64x64x3 |
|
# Convolutional layers |
