CAGAN: Consistent Adversarial Training Enhanced GANs

Authors: Yao Ni, Dandan Song, Xi Zhang, Hao Wu, Lejian Liao

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that our method can obtain state-of-the-art Inception scores of 9.17 and 10.02 on supervised CIFAR-10 and unsupervised STL10 image generation tasks, respectively, as well as achieve competitive semi-supervised classification results on several benchmarks. Importantly, we demonstrate that our method can maintain stability in training and alleviate mode collapse.
Researcher Affiliation Academia Yao Ni, Dandan Song, Xi Zhang, Hao Wu and Lejian Liao Lab of High Volume language Information Processing & Cloud Computing Beijing Lab of Intelligent Information Technology School of Computer Science & Technology, Beijing Institute of Technology {niyao, sdd, xi zhang, hao wu, liaolj}@bit.edu.cn
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access information (e.g., specific link, explicit statement of release in supplementary materials) for its source code.
Open Datasets Yes To investigate the effectiveness of our CAGAN on image generation task, we conduct experiments on two benchmark datasets: CIFAR-10 [Krizhevsky, 2009] and STL-10 [Coates et al., 2011]. CIFAR-10 contains 50,000 labeled training images of size 32 32 from 10 classes. ... STL-10 is subsampled from Image Net which is more diverse than CIFAR-10, and it contains 100,000 unlabeled images of size 96 96. ... MNIST, SVHN, and CIFAR-10.
Dataset Splits No The paper mentions using training images and an entire training set for unsupervised training but does not explicitly provide details about specific training/validation/test splits, percentages, or a defined validation set.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types) used for running its experiments.
Software Dependencies No The paper mentions the use of 'Adam optimizer' but does not specify any software dependencies with version numbers (e.g., programming language version, framework version, library versions).
Experiment Setup Yes We keep the hyper-parameters all the same on CIFAR-10 and STL-10 for all the experiments. In particular, we follow original WGAN-GP set λGP = 10, mini-batch size of 64 when training D, and mini-batch size of 128 when training G. We use Adam optimizer with a learning rate of 0.0002, β1 = 0, β2 = 0.9 to train G and D, and the learning rate is decreased linearly to 0. For consistent adversarial hyper-parameters, we set λf = 0.1, λCA = 2, and training totally 700 epochs.