In SupeRAuGAN we implement a novel data augmentation technique tailored to Generative Adversarial Networks in order to reduce discriminator overfitting and stabilize training. This technique was first described by Karras et al. and applied in an image generation from latent space task. We experiment such approach in a super resolution setting using a slightly modified SRGAN (described in Ledig et al.) achieving promising results when using a small amount of data.
More info about the whole project available at paper and slides
Original implementation available at repo
To get a local copy up and running follow these simple steps.
The project provide a Pipfile file that can be managed with pipenv.
pipenv installation is strongly encouraged in order to avoid dependency/reproducibility problems.
- pipenv
pip install pipenv- Clone the repo
git clone https://gitlab.com/reddeadrecovery/superaugan- Install Python dependencies
pipenv installHere's a brief description of each and every file in the repo:
model.py: Model definitiondata_utils.py: dataset loading utils and preprocessingtrain.py: GAN training filetest.py: GAN testing fileaugment.py: Data augmentation pipeline (taken from here)
Folders torch_utils and dnnlib are vendored dependencies of augment.py
Original downsampled images on the left, ground truthimages in the center and generated images on the right
Under the supervision of Leonardo Galteri
Visual and Multimedia Recognition © Course held by Professor Alberto Del Bimbo - Computer Engineering Master Degree @University of Florence