Improved Training of Generative Adversarial Networks Using Representative Features

Authors: Duhyeon Bang, Hyunjung Shim

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental evaluations show RFGAN effectiveness, improving existing GANs including those incorporating gradient penalty (Kodali et al., 2017; Gulrajani et al., 2017; Fedus et al., 2017). Section 4 summarizes the results of extensive experiments including simulated and real data. The quantitative and qualitative evaluations show that the proposed RFGAN simultaneously improved image quality and diversity.
Researcher Affiliation Academia Duhyeon Bang 1 Hyunjung Shim 1 1School of Integrated Technology, Yonsei University, South Korea. Correspondence to: Hyunjung Shim <kateshim@yonsei.ac.kr>.
Pseudocode No The paper describes the model and gradient updates mathematically but does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper references a GitHub link (https://github.com/carpedm20/DCGAN-tensorflow) for a baseline DCGAN implementation but does not state that the code for the proposed RFGAN methodology is publicly available or provide a link to it.
Open Datasets Yes For quantitative and qualitative evaluations, we include simulated and three real datasets: Celeb A (Liu et al., 2015), LSUN-bedroom (Yu et al., 2015), and CIFAR-10 (Krizhevsky & Hinton, 2009), normalizing between -1 and 1.
Dataset Splits No The paper mentions using training data and generating samples for evaluation but does not provide specific train/validation/test dataset splits (e.g., percentages or sample counts) or cross-validation details for reproducibility.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, memory, or other computer specifications used for running the experiments.
Software Dependencies No The paper refers to 'TensorFlow' implicitly through a baseline code link, but it does not provide specific version numbers for TensorFlow or any other software dependencies like programming languages or libraries used in their experiments.
Experiment Setup Yes input dimensionality is set at (64, 64, 3)... modify network dimensions for the CIFAR-10 dataset, fitting the input into (32, 32, 3)... drew 500 k images randomly from the LSUN bedroom dataset for efficient training and comparison. We use exactly the same hyper-parameters, metrics, and settings throughout this paper, as suggested for a baseline GAN... The WGAN-GP generator is updated once after the discriminator is updated five times. Following the reference code1, other networks are trained by updating the generator twice and the discriminator once.