A modified deep learning architecture for very fast image 2 image translation, based primarily off of NVIDIA's UNIT
Great question! The goal of the architecture is to take images that are in different domains such as summer/winter, zebra/horses and learn a mapping from one image to another. A full explanation of the architecture can be found here
-
Clone the repo:
https://github.com/BradleyBrown19/UNET-UNIT.git -
Open notebook Train.ipynb
-
Change path variable to location of dataset
Dataset should be organized by having a folder containing two directories called TrainA, TrainB
- Run the rest of the cells and watch the magic happen!
After 2 epochs of training, approximately 5k iterations
This architecture is based off of NVIDIA's UNIT and lots of inspiration was drawn from the fastai library.