photobashing-gan

This repository implements the PhotobashingGAN (or BashGAN for short), which was developed for a CSC420H5 (Introduction to Image Understanding at the University of Toronto) final project.

Dependencies

Please create and activate a conda environment with the dependencies listed in environment.yml. Please also install torchmetrics[image] via pip.

Download and unzip the 256.zip and eval.zip file stored in this Google Drive into the data/ folder, and download the generator weights into the results/ folder.

Repository Organization

The following summarizes the core files and folders of this repository:

• data/: the data folder.
  • 256/: the folder containing the landscape dataset (in 256/raw/) and corresponding color maps (in 256/cmaps/) on which the BashGAN was trained. We initially experimented with a downscaled 128x128 resolution dataset, but omitted the code and files from this repository.
  • extract_cmaps.py: preprocessing script used to extract the color maps from the landscape photos and cache the results.
• results/: the folder containing experimental results (plots, logs, etc.).
• src/: the source code folder.
  • cmap/: the folder containing utility methods used to extract a color map from an image. The entry method quantize_colors() can be found in cmap/quantize.py.
  • experimental/: the folder containing the training scripts used to train the BashGAN and compute or visualize experimental results.
    • train_bash_gan.py: script used to train BashGAN.
    • eval_bash_gan.py: script used to evaluate BashGAN (computing FID and LPIPS scores).
  • gan/: the folder containing the implementation of BashGAN. The actual high-level model can be found in gan/gan.py, while the other files implement the submodules (e.g. layers, generator, discriminator).

Reproducibility

The specific commands used to train and evaluate the GANs are given in the scripts/ folder.

References

• The landscape dataset used to train and evaluate BashGAN created by cleaning a subset of LHQ256, which was introduced by ALIS. The full original dataset (90k images) can be downloaded on their universome/alis repository.
• The BashGAN architecture was adapted from the SPADE architecture, found in the NVlabs/SPADE repository. We reimplement their code in a modified, simplified, and downscaled manner in gan/.
• The posterchild method of quantize_colors() uses the pallette extraction algorithm used by PosterChild. The files cmap/simplify_convexhull.py and cmap/trimesh.py are directly taken from the tedchao/PosterChild repository.