Stabilizing Training of Generative Adversarial Networks through Regularization

Authors: Kevin Roth, Aurelien Lucchi, Sebastian Nowozin, Thomas Hofmann

NeurIPS 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of this regularizer accross several architectures trained on common benchmark image generation tasks. Our regularization turns GAN models into reliable building blocks for deep learning. We demonstrate the stability of the regularized training procedure and to showcase the excellent quality of the samples generated from it, we trained various network architectures on the Celeb A [17], CIFAR-10 [15] and LSUN bedrooms [32] datasets.
Researcher Affiliation Collaboration Kevin Roth Department of Computer Science ETH Zürich kevin.roth@inf.ethz.ch Aurelien Lucchi Department of Computer Science ETH Zürich aurelien.lucchi@inf.ethz.ch Sebastian Nowozin Microsoft Research Cambridge, UK sebastian.Nowozin@microsoft.com Thomas Hofmann Department of Computer Science ETH Zürich thomas.hofmann@inf.ethz.ch
Pseudocode Yes Algorithm 1 Regularized JS-GAN. Default values: γ0 = 2.0, α = 0.01 (with annealing), γ = 0.1 (without annealing), n' = 1
Open Source Code Yes Code available at https://github.com/rothk/Stabilizing_GANs
Open Datasets Yes We trained various network architectures on the Celeb A [17], CIFAR-10 [15] and LSUN bedrooms [32] datasets.
Dataset Splits No The paper mentions using minibatches for training and specific datasets (Celeb A, CIFAR-10, LSUN bedrooms) but does not provide explicit details about the train/validation/test dataset splits, such as percentages or sample counts, or refer to standard splits with specific citations.
Hardware Specification No The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies No The paper mentions using the Adam optimizer and an open-source implementation from [10], but it does not provide specific version numbers for any software dependencies like programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes All networks were trained using the Adam optimizer [13] with learning rate 2 × 10−4 and hyperparameters recommended by [26]. We trained all datasets using batches of size 64, for a total of 200K generator iterations in the case of LSUN and 100k iterations on Celeb A. Algorithm 1 Regularized JS-GAN. Default values: γ0 = 2.0, α = 0.01 (with annealing), γ = 0.1 (without annealing), n' = 1