โญ Star this project on GitHub โ it helps!
Super-Resolution GAN is one of the best examples in GAN's application. It allows to imporove the quality of input low-resolution image. As Generator this GAN uses SRResNet (Super-Resolution Residual Network). In order to provide better quality of the output image the network is trained on three different loss functions.
You can evaluate pretrained Super-Resolution GAN on your images.
To do this use eval.py
.
The Super-Resolution GAN was trained on STL10 dataset from torchvision.dataset
.
Before to train both Generator and Discriminator we should pretrain our Ganarator on dataset with Pixel-wise Loss function.
See Super-Resolution [GAN WarmUp] for Generator's warmup.
After Generator warmup we train booth Generator and Discriminator with their loss functions. The Generator loss consists of Adverserial loss(BCE loss between fake prediction and target), Model Based loss(feature based MSE loss between fake and real images) and Pixel-wise loss(MSE loss between fake and real images).
After 100 epochs of training:
See Super-Resolution for SR-GAN's training.
This project is licensed under MIT.