Generating Pokémon with various GAN generator architectures, using pytorch. Part of the Seminar on Computer Vision, University of Helsinki, 2020.

I got Generative Adversarial Networks as the subject for my seminar report, and I decided to pit two advanced GAN generator architectures agains each other, using a tried-and-true DCNN as the baseline. The seminar report, with a bunch of images, explanation and results, is available here.

The code is still quite messy. I'll try to get around to cleaning it a bit soon!

1. Make sure you have Python 3 installed
2. Install requirements: pip3 install -r requirements.txt
3. Check that everything works: python3 ./discriminator.py
   • This should result in Trainable parameters: 2765568
4. You're good to go!

The raw data is expected to consist of

• .png -format images with
• one row of rectangular frames each

Personally, I used this dataset, and the folders in data.py are set up so that this dataset will work if extracted under 3d/3D Battlers in the project directory. Alternatively, you can change SPRITE_LOCATIONS to load data from elsewhere.

To generate a dataset for training, just make sure that you have valid sprites in SPRITE_LOCATIONS and run data.py. This will split the input sprites, pick a given number of frames from them and save the frames under data/pokemon (where PyTorch expects it to be).

To train the models, just comment out the relevant rows under if __name__ == '__main__' in training.py and run it. (I know, it's a bit of a mess, sorry!)

This repo includes some nice interactive visualizations for SAGAN and StyleGAN. To see those, run sagen_visualize.py or stylegen_visualize.py, respectively.