Taming GANs with Lookahead-Minmax

Authors: Tatjana Chavdarova, Matteo Pagliardini, Sebastian U Stich, François Fleuret, Martin Jaggi

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on MNIST, SVHN, CIFAR-10, and Image Net demonstrate a clear advantage of combining Lookahead minmax with Adam or extragradient, in terms of performance and improved stability, for negligible memory and computational cost.
Researcher Affiliation Academia Tatjana Chavdarova EPFL Mattéo Pagliardini EPFL Sebastian U. Stich EPFL François Fleuret University of Geneva Martin Jaggi EPFL
Pseudocode Yes Algorithm 1 General Lookahead Minmax pseudocode.
Open Source Code Yes Our source code is available: https://github.com/Chavdarova/LAGAN-Lookahead_Minimax.
Open Datasets Yes Datasets. We used the following image datasets: (i) MNIST (Lecun & Cortes, 1998), (ii) CIFAR-10 (Krizhevsky, 2009, 3), (iii) SVHN (Netzer et al., 2011), and (iv) Image Net ILSVRC 2012 (Russakovsky et al., 2015), using resolution of 28 28 for MNIST, and 3 32 32 for the rest.
Dataset Splits No The paper mentions training and evaluating on standard datasets but does not explicitly provide details about train/validation/test splits (e.g., percentages or sample counts) within the main text or appendices.
Hardware Specification No The paper mentions 'common computational resources' but does not provide specific hardware details such as GPU/CPU models or memory specifications used for the experiments.
Software Dependencies No For our experiments, we used the Py Torch2 deep learning framework. For experiments on CIFAR-10 and SVHN, we compute the FID and IS metrics using the provided implementations in Tensorflow3 for consistency with related works. The full unrolling that performs the backpropagation on the unrolled discriminator was implemented using the Higher7 library. (No version numbers are provided for these software packages.)
Experiment Setup Yes Table 4: List of hyperparameters used in Figure 4. η denotes the learning rate, β1 is defined in equation 5, and α and k in Alg. 1. Tables 8, 9, 10, 11, 12 provide detailed hyperparameters such as ηG, ηD, Adam β1, Batch-size, Update ratio r, Lookahead k, Lookahead α, and Unrolling steps for experiments on MNIST, SVHN, CIFAR-10, and ImageNet.