Let’s try to eliminate the job of switching male shot into feminine and vice versa. To do this we’d like datasets with female and male photos. Nicely, CelebA dataset is perfect for our desires. Actually intended for cost-free, it’s got 200k pictures and 40 digital labeling like sex, glasses, Putting oncap, BlondeHair, etc.
This dataset have 90k images of male and 110k female images. That’s efficiently enough for the DomainX and DomainY. The typical size of face on these imagery is not large, only 150×150 pixels. So we resized all taken faces to 128×128, while keeping the aspect relation and making use of black colored qualities for graphics. Very common enter to the Cycle-GAN could appear to be this: