Adversarial Lipschitz Regularization

Authors: Dávid Terjék

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Applying ALP on the critic in WGAN (WGAN-ALP), we show state-of-the-art performance in terms of Inception Score and Fréchet Inception Distance among non-progressive growing methods trained on CIFAR-10, and competitive performance in the high-dimensional setting when applied to the critic in Progressive Growing GAN trained on Celeb A-HQ. To evaluate the performance of WGAN-ALP, we trained one on CIFAR-10, consisting of 32 32 RGB images, using the residual architecture from Gulrajani et al. (2017), implemented in Tensor Flow.
Researcher Affiliation Industry Dávid Terjék Robert Bosch Kft. Budapest, Hungary david.terjek@hu.bosch.com
Pseudocode No The paper describes the methods through mathematical formulations and textual explanations but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Source code to reproduce the presented experiments is available at https://github.com/dterjek/adversarial_lipschitz_regularization.
Open Datasets Yes To evaluate the performance of WGAN-ALP, we trained one on CIFAR-10, consisting of 32 32 RGB images... To show that ALR works in a high-dimensional setting as well, we trained a Progressive GAN on the Celeb A-HQ dataset (Karras et al., 2018), consisting of 1024 1024 RGB images.
Dataset Splits No No explicit mention of a dedicated validation dataset split with specific sizes or percentages for the main GAN experiments in Section 4. While Section A.1 mentions a validation split for a semi-supervised learning experiment, it is not for the primary focus of the paper.
Hardware Specification No The paper does not provide specific details on the hardware used for running experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions "Tensor Flow" and "Py Torch" but does not specify their version numbers or any other software dependencies with version details.
Experiment Setup Yes Closely following Gulrajani et al. (2017), we used the Adam optimizer (Kingma and Ba, 2015) with parameters β1 = 0, β2 = 0.9 and an initial learning rate of 2 10 4 decaying linearly to 0 over 100000 iterations, training the critic for 5 steps and the generator for 1 per iteration with minibatches of size 64 (doubled for the generator). We used (17) as a loss function to optimize the critic. K = 1 was an obvious choice, and we found λ = 100 to be optimal... The hyperparameters of the approximation of radv were set to ξ = 10, Pϵ being the uniform distribution over [0.1, 10], and k = 1 power iteration.