BYO-GAN

As a part of my independent study for Junior year, my final project was to recreate StyleGan in Python using Pytorch. I used some notes from a Coursera course I took, the original whitepapers, and looked at how others had implemented the same architecture in order to better understand how StyleGan worked.

Whitepapers:

• A Style-Based Generator Architecture for Generative Adversarial Networks
  • Official Implementation
• Progressive Growing of GANs for Improved Quality, Stability, and Variation
  • Official Implementation

Datasets:

• Anime Profile Dataset
  • License
• Abstract Art Dataset
• Abstract Art Images
• FFHQ Dataset

MISC:

Big thanks to these projects for helping me troubleshoot issues I had!

• huangzh13
• rosinality

Thanks to Samuel Prevost for directly helping me troubleshoot issues.

How to run

Preparing a dataset

Create a folder under /data, name it however you want.

Place all of your dataset images into this new folder. It will work best if they are all 512x512.

Run python prep.py [path to images] [start size (4)] [end size (512)].

This script is SUPER dodgy and thrown together, so be prepared to tweak it in order to make it work. It essentially moves those original images into a new /data/[name]/original/images folder. Then, it resizes every image to match progressive growth and saves it under separate datasets under /data/[name]/prepared, where name refers to the original folder (e.g. /data/art).

Training

Edit the config.txt file and create a configuration setting to your liking. Use the two examples as a template. You can override any key, but do NOT delete anything under the DEFAULT setting.

python main.py [config name] -c checkpoint.pth

For instance,

python main.py abstract-art

runs with the abstract-art configuration.

Getting Full Resolution Samples

python generate_samples.py ./checkpoints/checkpoint.pth 64 -o ./output/ -d cpu 

will generate 64 images from the model saved at ./checkpoints/checkpoint.pth and save them in the ./output folder. It will use the CPU.

python generate_samples.py -h

for more info.

Results

As my resources are limited, I was unable to run my program to completion. See Nvidia's implementation for higher fidelity Results.

However, for the time that I did run this program, I did get meaningful results that show that this implementation is indeed functional.

16 FFHQ images:

16 Abstract Art images: