Intuition on how to construct your GANs

So far for this course, I have been relying on the instructions on how do I actually finish my assignments. But I envision myself working in developing GANs, it made me think: How does one develop intuition for the components needed to be included in the GAN? Some questions I wanted to inquire: How much Conv2d layers there needed, why do we not need to pool, how many layers does we need, and how does the architecture affect the performance of your GAN by intuition? Is there a prerequisite I missed? Imagining myself taking a GAN job, and I was asked to create one from scratch. I don’t think I’ll be able to do it without referring to a ready made example or without any guidance, I think I’ll be lost. Would you instructors happen to have any input or any share of experience whether you can create GANs without any reference? Or would cloning a GAN from GitHub be sufficient when it comes to working with the AI industry?

The whole process of AI is based on previous experiences, as Einstein said I stand on the shoulder of giants, it is the case here too. You start with the known and maybe by trial and error (there are search techniques to test a variety of model architectures and choose a better one like for ex. grid search) but fundamentally are based on trial and error.

Try to learn the principles of working and the rest with enough will can take over. Pooling layers help pool information because from cnov2d layers information is distilled its good to pool what ever is left to have a s much as possible.