Using GANs to generate pokemons
A custom dataset was used by scraping DuckDuckGo image search. A python script for scraping images off DuckDuckGo is provided. The script is modified from duckduckgo-images-api. A dataset of 2109 fire Pokemons was gathered. All the images were augmented to obtain a final dataset of size 37962 images with each image rescaled to 96x96.
This project uses a Wasserstein GAN with Gradient Penalty to generate images of new Pokemons. The training was carried out on Google Colab. A Colab notebook is provided in the repository.
The training was carried out for about 12000 iterations, which translates to 10 epochs at a batch size of 32. The generated images are not yet of a very good quality.