Understanding and Stabilizing GANs’ Training Dynamics Using Control Theory

Authors: Kun Xu, Chongxuan Li, Jun Zhu, Bo Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results show that our method can effectively stabilize the training and obtain state-of-the-art performance on data generation tasks.
Researcher Affiliation Collaboration 1Dept. of Comp. Sci. & Tech., Institute for AI, BNRist Center, Tsinghua-Bosch ML Center, THBI Lab, Tsinghua University, Beijing, China.
Pseudocode Yes Algorithm 1 Cloosed-loop Control GAN
Open Source Code No Our code is provided HERE. (Note: 'HERE' is a placeholder and not an actual link to a repository)
Open Datasets Yes We now empirically verify our method on the widely-adopted CIFAR10 (Krizhevsky et al., 2009) and Celeb A (Liu et al., 2015) datasets.
Dataset Splits No The paper mentions using CIFAR10 and Celeb A datasets but does not explicitly provide details about the validation split (e.g., percentages, counts, or specific references to predefined validation sets used in their experiments).
Hardware Specification Yes For instance, our method can conduct approximate 8 iterations per second of training on Celeb A whereas Reg-GAN can only conduct 4 iterations per second on Geforce 1080Ti.
Software Dependencies No The paper mentions architectures like ResNet and ReLU activation, but it does not provide specific version numbers for any software dependencies (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes The batch size is 64, and the buffer size Nb is set to be 100 times of the batch size for all settings. We manually select the coefficient λ among {1, 2, 5, 10, 15, 20} in Reg-GAN s setting and among {0.05, 0.1, 0.2, 0.5} in SN-GAN s setting.