Skip to content

Implementation of SRGAN trained on an anime faces dataset

Notifications You must be signed in to change notification settings

eyess-glitch/Super-resolution-GAN

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

13 Commits
 
 
 
 
 
 
 
 

Repository files navigation

1. Overview

SRGAN is a neural network architecure firstly introduced in the following paper for upscaling images.

2. Training process

This repository contains the source code for the implementation of the SRGAN discussed in the following paper link. The model was trained in the following way : firstly, the generator is trained with a non GAN approach by minimizing the MSE loss function between the upscaled image and reconstructed image for roughy 40.000 iteration of 32 batches each on a Google Colab notebook (free-tier); next, both discriminator and generator are trained with a GAN approach for approximately 20.000 iterations on Google Colab (Pro version).

3. Dataset used for training

The dataset (available at the following kaggle link) on which the model was trained is organized in the following way : a folder containing various subfolders, each associated with an id, containing the original image and the downsampled image. The downsampled image was obtained by executing the scripting downsampling.py, which applies a 4x downsampling factor to a specific image. The organization of the folders in different subfolders was realized through a bash script.

4. Model's weights

The model's weights are available here.