Drawing system vs. trained discriminator

Hi, I’m doing part 1 these days. I’m trying to build my own project idea.
I would appreciate any advice on how to go about implementing it - general directions, examples, or tips all welcome.

It’s kind of (but not quite) a GAN. On the generator side, I have a drawing system that produces images. As a simple example, let’s say it’s drawing eight circles, and can vary their location and size.
On the discriminator side, would be a trained image classifier (eg, Resnet) that would evaluate how much the drawing resembles an apple, therefore providing a loss for the drawing system to optimize for (hopefully reaching a 1.0 predicted apple).
Here’s a little “schematic” of what I’m trying to do:

So a little bit like a GAN, but the discriminator is already well-trained. There is a similar idea from Tom White, “Perception Engines” but not much explanation on how it’s done over there:



Really interesting idea – looking forward to seeing more :slight_smile:
Check out this related idea, maybe there is something helpful here :slight_smile:


1 Like