Source code of our paper Taming GANs with Lookahead–Minmax, ICLR 2021. Equal contribution with Matteo Pagliardini, and joint work with Sebastian Stich, François Fleuret, and Martin Jaggi.
To tackle the known challenges of minmax optimization of: (i) rotational joint vector field and (ii) sensitivity to noise induced by the stochastic gradient, we propose the Lookahead-Minmax algorithm. It consists of periodically taking an iterate that lies on a convex combination of the current and a past iterate.
The subdirectories contain:
lagan: the experiements on CIFAR10, SVHN and ImageNet.mnist: the experiement on MNIST.bilinear_game: the Colab-notebooks to reproduce the results on batch and stochastic bilinear game.
See also: paper, poster, and slides.
@inproceedings{chavdarova2021taming,
title={{Taming GANs with Lookahead-Minmax}},
author={Tatjana Chavdarova and Matteo Pagliardini and Sebastian U Stich and Fran{\c{c}}ois Fleuret and Martin Jaggi},
booktitle={International Conference on Learning Representations},
year={2021}
}