You heard of gans? GANS. Sounds like guns, cool.
but wait, there is another.
As you can probably see, I am prepearing my own private library for automated GAN training. It may be a part of some bigger project, or i will train a GAN that generates random memes. Who cares.
or whatever it is called there, is an interesting type of machine learning model, as it actually consists of two models. These are professionally called generator and discriminator. But I prefer to call them
as they resamble them mostly.
Artist goal is to create some random art. But not completly random, he has a topic. Lets say he's goal is to paint cute anime girls.
But he is actually colorblind so he only uses gray.
(Btw he is not colorblind anymore, he understands colors now)
But oh wait, our Artist never actually painted anime girl. He actually never painted anything. That is why we have our Critique. Critique has an access to database of pictures of cute anime girls [fcking weeboo] and he nows his shit. If he will get some ugly anime painting he will throw her into the thrash. That is why our Artist will try to create anime girl that will be indistinguishable from other Anime girls. Our Critique must be good enough to istinguish true anime girl from fake one.
some block diagram for no purpose:
So yeah, that's how it works. For some mathematicians here: I am currently experimenting with different loss functions like wasserstien discriminator loss function to minimize the mode collapse problem. As it should be a fully automated process of training I want to find such hyperparameters that 99,99% of time my GAN will be trained create at least nice approximation of the dataset.
Thats all, bye.
https://www.tensorflow.org/tutorials/generative/dcgan
https://www.kaggle.com/splcher/animefacedataset
https://developers.google.com/machine-learning/gan/loss