/GAN_Convergence

A quick review and experiment of the (lack of) theoretical convergence guarantees of Generative Adverserial Networks. Then the Wasserstein Distance as remediation.

Primary LanguageJupyter Notebook

GAN_Convergence

This notebook contains an experiment to empirically compare the convergence of a standard GAN with that of a Wasserstein GAN of identical architecture (with the exception of outputs). The models attempt to generate greyscale images utilizing the CelebA dataset.

Installing

Install the necessary packages

pip3 install -r requirements.txt

TODO

  • Citations for CelebA dataset
  • Embed some images of the generated faces
  • Clean up directory structure
  • Split the actual model into a .py file instead of a class within the notebook
  • Re-edit the latex report and upload it
  • Improve project description