README.md

Super Resolution using GANs

• SRGAN is an image-to-image translation model using Generative Adversarial Networks
• This project is developed with DVC pipeline
• Dataset used - DIV2K - bicubic downscaling x8 competition
• Model is only trained for 35 epochs on Google Colab
• You can train it according to your need by changing parameters in params.yaml file
• Model Architecture:

Requirements

• Create a new virtual environment.
• Python 3.6 or greater.
• DVC
• Run Command-

pip install -r requirements.txt

Usage:

• Change parameters and directories according to your system (recommendation - do not change) in params.yaml.
• Setup the Directory named Results in the root directory.
• Make sure you have DVC initialized in the root directory
  You can do it with the command:

dvc init

• Add data\train & data\val for data tracking in dvc by using command: For Windows-

dvc add data\train\HR
dvc add data\train\LR
dvc add data\validation\HR
dvc add data\validation\LR

For Linux-

dvc add data/train/HR
dvc add data/train/LR
dvc add data/validation/HR
dvc add data/validation/LR

• Finally, run:

dvc repro

• Trained models will be saved in saved_models directory.
• Results will be saved in Results\present_datatime directory with name present_time_result.png.

Todo:

• Integrate mlflow for better tracking
• Create webapp
• Deployment

References:

• Reference of Subpixel_conv2D layer and SubpixelConv2D
• Reference of loss functions
• Blog

Warning: Do not try to change dvc.lock file and .dvc directory