As we know, generally speaking, there are two models in GANs.
Generator model
Discriminator model
My point of doubt is that:
Why people do NOT use Generator model for the training set augmentation, even though the Generator model generates very realistic samples?
Having said that, GANs are difficult and slow to train, and you should first pursue traditional data augmentation where your relative gains will be significant.
In a study group I was in, we decided to use GANs to augment the tiny dataset provided (one of the better tricks among all we tried, ~1-2% boost) in a contest in analyticsvidya at that time. Members of the group were placed 1st, 2nd, 3rd, 4th, 5th, 9th, 11th, 13th and 19th.
GANs generated quite a few rounds of drinks for us (we got all the cash prizes).