Demostration examples from Cumulant GAN paper.

• Code for all demonstrations can be found in Dip's repo.
• TensorFlow 2 implementation.

Python, NumPy, TensorFlow 2, SciPy, Matplotlib

Line defined by β + γ = 1 has a point symmetry. The central point, (0.5,0.5) corresponds to the Hellinger distance. For each point, (α, 1 − α), there is a symmetric one, i.e., (1 − α, α), which has the same distance from the symmetry point. The respective divergences have reciprocal probability ratios (e.g., KLD & reverse KLD, χ^2-divergence & reverse χ^2-divergence, etc.). Each point on the ray starting at the origin and pass through the point (α, 1 − α) also corresponds to (scaled) Rényi divergence of order α. These half-lines are called d-rays.

KLD minimization that corresponds to (β, γ) = (0, 1) tends to cover all modes while reverse KLD that corresponds to (β, γ) = (1, 0) tends to select a subset of them. Hellinger distance minimization produces samples with statistics that lie between KLD and reverse KLD minimization while Wasserstein distance minimization has a less controlled behavior.

Wasserstein Distance (β, γ) = (0, 0)	Kullback-Leibler Divergence (β, γ) = (0, 1)	Reverse Kullback-Leibler Divergence (β, γ) = (1, 0)	-4log(1-Hellinger^2) (β, γ) = (0.5, 0.5)

Wasserstein Distance (β, γ) = (0, 0)	Kullback-Leibler Divergence (β, γ) = (0, 1)	Reverse Kullback-Leibler Divergence (β, γ) = (1, 0)	-4log(1-Hellinger^2) (β, γ) = (0.5, 0.5)

Wasserstein Distance (β, γ) = (0, 0)	Kullback-Leibler Divergence (β, γ) = (0, 1)	Reverse Kullback-Leibler Divergence (β, γ) = (1, 0)	-4log(1-Hellinger^2) (β, γ) = (0.5, 0.5)

To reproduce the achieved results run the main.py script with the corresponding arguments. The output is a directory with some checkpoints during the traininng proccess, the corresponding plots and finally a gif that visualizes the whole training.

@ARTICLE{9750393,
  author={Pantazis Yannis, Paul Dipjyoti, Fasoulakis Michail, Stylianou Yannis and Katsoulakis Markos A.},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  title={Cumulant GAN},
  year={2022},
  pages={1-12},
  doi={10.1109/TNNLS.2022.3161127}}