Why Spectral Normalization Stabilizes GANs: Analysis and Improvements

Authors: Zinan Lin, Vyas Sekar, Giulia Fanti

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

Reproducibility Variable Result LLM Response
Research Type Experimental Theoretically, we show that BSSN gives better gradient control than SN. Empirically, we demonstrate that it outperforms SN in sample quality and training stability on several benchmark datasets.
Researcher Affiliation Academia Zinan Lin Carnegie Mellon University Pittsburgh, PA 15213 zinanl@andrew.cmu.edu Vyas Sekar Carnegie Mellon University Pittsburgh, PA 15213 vsekar@andrew.cmu.edu Giulia Fanti Carnegie Mellon University Pittsburgh, PA 15213 gfanti@andrew.cmu.edu
Pseudocode No The paper describes procedures but does not include a clearly labeled 'Pseudocode' or 'Algorithm' block or figure.
Open Source Code Yes The code for reproducing the results is at https://github.com/fjxmlzn/BSN.
Open Datasets Yes More specifically, we conducts experiments on CIFAR10, STL10, Celeb A, and Image Net (ILSVRC2012)
Dataset Splits Yes All experimental details are attached in Apps. N to S.
Hardware Specification Yes Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See App. T.
Software Dependencies No The paper refers to its open-source code repository for reproducibility details ('See https://github.com/fjxmlzn/BSN'), but the provided text does not explicitly list specific software dependencies with version numbers.
Experiment Setup Yes All experimental details are attached in Apps. N to S.