RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks

Authors: Divya Saxena, Jiannong Cao, Jiahao Xu, Tarun Kulshrestha

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our results consistently demonstrate RG-GAN s robust performance across a variety of scenarios, including different GAN architectures, datasets, and degrees of data scarcity, reinforcing its value as a generic training methodology.
Researcher Affiliation Academia Divya Saxena1, Jiannong Cao1, Jiahao Xu2, Tarun Kulshrestha3 1Department of Computing, The Hong Kong Polytechnic University, Hong Kong 2Department of Computer Science and Engineering, University of Nevada, Reno, USA 3University Research Facility of Big Data Analytics, The Hong Kong Polytechnic University, Hong Kong
Pseudocode Yes Algorithm 1: RG-GAN training
Open Source Code Yes Code can be found at this link: https://github.com/Intellicent AI-Lab/RG-GAN
Open Datasets Yes Datasets: RG-GAN s effectiveness is exhaustively tested through evaluation on a wide spectrum of datasets, encompassing various resolutions 32 32, 64 64, 256 256 and 1024 1024. These datasets range from the commonly used CIFAR-10 (32 32) and Tiny-Image Net (64 64) (on 10%, 20%, 50% and 100% of training dataset), to specific animal face (AF) datasets (256 256)(Si and Zhu 2011) (Dog (389 images) and Cat (160 images)), 100-shot datasets (256 256) (Zhao et al. 2020) (Obama, Panda, and Grumpy-Cat), and the varying size of high-resolution FFHQ (1024 1024) (Karras, Laine, and Aila 2019).
Dataset Splits No The paper specifies percentages of the training dataset used (e.g., '10%, 20%, 50% and 100% of training dataset') and mentions using 'test data' as a reference for FID calculation. However, it does not explicitly define or refer to a distinct 'validation' dataset split for hyperparameter tuning or early stopping.
Hardware Specification Yes For 1024 1024 resolution, experiments are performed on NVIDIA RTX A6000 GPU (48GB), while for remaining resolutions, NVIDIA RTX 3090 GPU (24GB) is used.
Software Dependencies No The paper describes the base models and loss functions used (e.g., Pro GAN, Style GAN2, WGAN-GP), but does not specify software versions for libraries, frameworks, or programming languages (e.g., PyTorch version, Python version, CUDA version).
Experiment Setup Yes Augmented regeneration is applied when current epoch s FID is greater than the weighted average FID value calculated over the previous n (n=5 in all experiments) epochs. This increase in the FID value indicates a decline in the quality of the generated images, suggesting the network may have pruned some important connections. We set sparsity penalty (λ) 5e-12 for all experiments.