Generative Adversarial Training with DUGAN

[alexander-huang-commits]

SUMMARY:
DUGAN is an adversarial training network based off of https://arxiv.org/abs/2108.10772. The training script is written in a way to allow for different neural network architectures and data augmentation techniques (in DUGAN's case, it is cutmix) to be directly be integrated without rewriting the structure of the scripts.

CHANGES:
-There are only new files no deleted files
-Added Training script for DUGAN
-Added Cutmix data augmentation
-Added Unet with fixed convolution before input
-Added REDCNN https://arxiv.org/pdf/1702.00288
-Added modifications of LPD and ItNet

[AnderBiguri]

Thanks a lot! I may do some commits to this branch to change some names and relocate some files, mostly cosmetic, is that ok?

Also: is LPD_SIRT basically the standard LPD?

[alexander-huang-commits]

Thanks a lot! I may do some commits to this branch to change some names and relocate some files, mostly cosmetic, is that ok?

Also: is LPD_SIRT basically the standard LPD?

Yes, feel free to rearrange or edit any of the files! I have a separate local save of this work so don't worry about it.

Also, yes LPD_SIRT is just LPD initialized with SIRT using 50 iterations instead of FBP.