/Colorize

Conditional GAN for colorizing grayscale images

Primary LanguagePythonMIT LicenseMIT

Colorize

Conditional GAN for colorizing grayscale images. Uses UNet architecture for generator and Patch-Discriminator for discriminator.


Getting Started

  • Clone this repository
  • cd Colorize
  • Download the 2014 COCO validation dataset and extract it to data/images
  • Run python data/split.py to split the dataset into training and validation sets

How to Use

  • Change the hyperparameters in config.py if needed
  • (Optional) Run python pretrain_generator.py to pretrain the generator (L1 Loss)
  • If using pretrained generator, change the pretrained_gen parameter in config.py to true otherwise false
  • Run python train_GAN.py to train and save the model
  • Store the images to be colorized in ip folder
  • Update inference.py with the path to the checkpoint
  • Run python inference.py to colorize the images