Data Augmentation

[FranzHahn]

In your data augmentation you create three alternative version:

• flipped
• rotated left
• rotated right

I would suggest you also use flipped + rotated left and flipped + rotated right. Otherwise you are biasing your data, such that the faces more likely look in the direction that the original image faces.

Additionally, this should alleviate some of the zig-zagging that your losses are doing, as currently the discriminator favors faces that look in a specific way.

[zhj131]

I can't download the training data from the given links, can you send me a copy? my email: [email protected], thank you very much!