Consistency Regularization for Generative Adversarial Networks

Authors: Han Zhang, Zizhao Zhang, Augustus Odena, Honglak Lee

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

Reproducibility Variable Result LLM Response
Research Type Experimental This section validates our proposed CR-GAN method. First we conduct a large scale study to compare consistency regularization to existing GAN regularization techniques (Kodali et al., 2017; Gulrajani et al., 2017; Roth et al., 2017) for several GAN architectures, loss functions and other hyperparameter settings. We then apply consistency regularization to a state-of-the-art GAN model (Brock et al., 2018) and demonstrate performance improvement. Finally, we conduct ablation studies to investigate the importance of various design choices and hyper-parameters.
Researcher Affiliation Industry Han Zhang, Zizhao Zhang, Augustus Odena, Honglak Lee Google Research {zhanghan,zizhaoz,augustusodena,honglak}@google.com
Pseudocode Yes Algorithm 1 Consistency Regularized GAN (CR-GAN). We use λ = 10 by default.
Open Source Code Yes All our experiments are based on the open-source code from Compare GAN (Kurach et al., 2019), which is available at https://github.com/google/compare_gan.
Open Datasets Yes We validate our proposed method on three datasets: CIFAR-10 (Krizhevsky, 2009), CELEBA-HQ-128 (Karras et al., 2018), and Image Net-2012 (Russakovsky et al., 2015).
Dataset Splits No The paper specifies training and testing set sizes for CIFAR-10 (50K training, 10K testing) and Celeb A (27K training, 3K testing), and mentions Image Net-2012, but does not explicitly define a separate validation dataset split or state a methodology for hyperparameter tuning that uses a validation set.
Hardware Specification Yes Here we show the actual training speed of discriminator updates for SNDCGAN on CIFAR-10 with NVIDIA Tesla V100.
Software Dependencies No The paper mentions using the 'Adam optimizer' and that experiments are 'based on the open-source code from Compare GAN', but does not provide specific version numbers for software libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages used.
Experiment Setup Yes For optimization, we use the Adam optimizer with batch size of 64 for all our experiments. We stop training after 200k generator update steps for CIFAR-10 and 100k steps for Celeb A. Appendix A (Table A1) provides hyper-parameters of the optimizer including learning rate (lr), Adam β1 and β2, and number of discriminator iterations per generator update (NDis) across 7 different settings.