Skip to content

akanametov/super-resolution

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

46 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Super-Resolution GAN

⭐ Star this project on GitHub — 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.

Table of content

Evaluation

You can evaluate pretrained Super-Resolution GAN on your images. To do this use eval.py.

Training

Database

The Super-Resolution GAN was trained on STL10 dataset from torchvision.dataset.

WarmUp of Generator

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.

Training with Discriminator

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).

Generator loss

Discriminator loss

After 100 epochs of training:

See Super-Resolution for SR-GAN's training.

License

This project is licensed under MIT.

Links