The first thing that I'd like to poiny out to is that the time required to make the following system can be much lower.
But I like to maintain an uppercap to allow room for client corrections.
A little about me:
I'm Arghya, a reseach student at IIIT Hyderabad.
My majors is in Computer Science and I am a Computational Linguist.
I've worked with GANs before for similar tasks and the same is something I'm comfortable with.
I'm also a mentor for Google CodeIn. I cam forward my resume is required.
A little Idea of my Ideas for the project:
Although GAN has shown great success in the realistic image generation, the training is not easy, The process is known to be slow and unstable.
Things I plan to do to achieve better and faster convergence:
(1) Feature Matching
Feature matching suggests to optimize the discriminator to inspect whether the generator’s output matches expected statistics of the real samples
(2) Minibatch Discrimination
With minibatch discrimination, the discriminator is able to digest the relationship between training data points in one batch, instead of processing each point independently.
(3) Historical Averaging
(4) One-sided Label Smoothing
When feeding the discriminator, instead of providing 1 and 0 labels, use soften values such as 0.9 and 0.1. It is shown to reduce the networks’ vulnerability.
(5) Virtual Batch Normalization (VBN)
Hope to work with you soon!