基于梯度范数差值的一种正则化方法
A Regularization Method Based on Gradient Norm Difference
摘要: 生成对抗网络(GANs)在学习从给定数据集指定的分布中采样方面非常成功，特别是给定数据集的数据量远大于其维度时。当数据有限时，经典的生成对抗网络生成的图像的质量会有显著降低，而输出正则化、数据增强、使用预训练模型和修剪等策略已被证明可以改善这种情况。然而这些方法常受限于特定的设置，例如预训练模型受限于数据的类型等。相比之下，本文提出的正则化方法通过优化鉴别器在真实图像与生成样本的梯度范数之间的差值来增强现有的生成对抗网络，并且具有很强的兼容性，适用于大多数现有的生成对抗网络。在数据有限的情况下显著的改善了训练成果。
Abstract: Generative adversarial networks (GANs) are very successful at learning to sample from a specified distribution of a given dataset, especially when the amount of data in a given dataset is much larger than its dimensions. Classical generative adversarial networks struggle when data is limited, while strategies such as output regularization, data augmentation, using pre-trained models, and pruning have been shown to bring improvements. However, these methods are often limited by specific set-tings. For example, pre-trained models are limited by the type of data. In contrast, the regulariza-tion method proposed in this paper enhances the existing generative adversarial network by opti-mizing the difference between the discriminator between the real image and the gradient norm of the generated sample, and has strong compatibility applicable to most existing generative adver-sarial networks. Training outcomes were significantly improved when data were limited.
文章引用：吴天宝, 徐芳, 张云轩. 基于梯度范数差值的一种正则化方法[J]. 应用数学进展, 2023, 12(3): 1367-1373. https://doi.org/10.12677/AAM.2023.123139

参考文献

[1] Goodfellow, I., Pouget-Abadie, J., Mirza, M., et al. (2014) Generative Adversarial Nets. NIPS’14: Proceedings of the 27th International Conference on Neural Information Processing Systems, Volume 2, 2672-2680.
[2] Karras, T., Aittala, M., Laine, S., et al. (2021) Alias-Free Generative Adversarial Networks. Advances in Neural Information Pro-cessing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, 6-14 Decem-ber 2021, 214-233.
[3] Wang, Z.W., She, Q. and Ward, T.E. (2021) Generative Adversarial Networks in Computer Vision: A Survey and Taxonomy. ACM Computing Surveys (CSUR), 54, 1-38. [Google Scholar] [CrossRef] 
[4] Wang, L., Ho, Y.-S. and Yoon, K.-J. (2019) Event-Based High Dynamic Range Image and Very High Frame Rate Video Generation Using Conditional Generative Adversarial Networks. Pro-ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, 15-20 June 2019, 10081-10090. [Google Scholar] [CrossRef] 
[5] Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O. and Cohen-Or, D. (2021) Designing an Encoder for StyleGAN Image Manipulation. ACM Transactions on Graphics, 40, 1-14. [Google Scholar] [CrossRef] 
[6] Liu, H.Y., Wan, Z.Y., Huang, W., Song, Y.B., Han, X.T. and Liao, J. (2021) Pd-GAN: Probabilistic Diverse GAN for Image in Painting. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, 20-25 June 2021, 9371-9381.
[7] Mao, X., Li, Q., Xie, H., et al. (2017) Least Squares Generative Adversarial Networks. Proceedings of the IEEE International Conference on Computer Vision, 2794-2802.
[8] Choi, J., Kim, T. and Kim, C. (2019) Self-Ensembling with GAN-Based Data Augmentation for Domain Adaptation in Semantic Segmentation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, 27 October-2 November 2019, 6830-6840. [Google Scholar] [CrossRef] 
[9] Zhu, J.-Y., Park, T., Isola, P. and Efros, A.A. (2017) Unpaired Image-to-Image Translation Using Cycle-Consistent Adver-sarial Networks. 2017 IEEE International Conference on Computer Vision (ICCV), Venice, 22-29 October 2017, 2242-2251. [Google Scholar] [CrossRef] 
[10] Zhuang, P., Koyejo, O. and Schwing, A.G. (2021) Enjoy Your Editing: Controllable GANs for Image Editing via Latent Space Navigation. International Conference on Learning Representations, ICLR 2021, Vienna, 4 May 2021, 124-136. [Google Scholar] [CrossRef] 
[11] Kumari, N., Zhang, R., Shechtman, E. and Zhu, J.-Y. (2022) Ensembling Off-the-Shelf Models for GAN Training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, 18-24 June 2022, 10641-10652.
[12] Chen, T., Cheng, Y., Gan, Z., Liu, J. and Wang, Z. (2021) Data-Efficient GAN Training beyond (Just) Augmentations: A Lottery Ticket Perspective. 35th Con-ference on Neural Information Processing Systems (NeurIPS 2021), 6-14 December 2021, 256-263.
[13] Zhang, H., Zhang, Z., Odena, A. and Lee, H. (2020) Consistency Regularization for Generative Adversarial Networks. International Conference on Learning Representations, Addis Ababa, 26-30 April 2020, 21655-21667.
[14] Tseng, H.-Y., Jiang, L., Liu, C., Yang, M.-H. and Yang, W. (2021) Regularizing Generative Adversarial Networks under Limited Data. Pro-ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, 20-25 June 2021, 7921-7931. [Google Scholar] [CrossRef] 
[15] Karras, T., Laine, S. and Aila, T. (2019) A Style-Based Generator Architecture for Generative Adversarial Networks. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, 15-20 June 2019, 4401-4410. [Google Scholar] [CrossRef] 
[16] Ba, J.L., Kiros, J.R. and Hinton, G.E. (2016) Layer Normalization. Preprint. arXiv: 1607.06450.
[17] Xiao, C., Zhong, P. and Zheng, C. (2018) Bourgan: Generative Networks with Metric Embeddings. NeurIPS 2018, Montreal, 3-8 December 2018, 136-152.
[18] Yazıcı, Y., Foo, C.-S., Winkler, S., Yap, K.-H., Piliouras, G. and Chandrasekhar, V. (2019) The Unusual Effectiveness of Averaging in GAN Training. 7th In-ternational Conference on Learning Representations, ICLR 2019, New Orleans, 6-9 May 2019, 25-31.
[19] Roth, K., Lucchi, A., Nowozin, S. and Hofmann, T. (2017) Stabilizing Training of Generative Adversarial Networks through Reg-ularization. Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Pro-cessing Systems 2017, Long Beach, 4-9 December 2017, 30-36.
[20] Xie, T., Du, Y.M., Wang, T., et al. (2020) Struc-ture Improved Conditional Spectral Normalization Generative Adversarial Networks for Image Recognition. 2020 Chinese Automation Congress (CAC), 842-846.
[21] Brock, A., Donahue, J. and Simonyan, K. (2018) Large Scale GAN Training for High Fidelity Natural Image Synthesis.
[22] Jiang, L., Dai, B., Wu, W. and Change Loy, C. (2021) Deceive d: Adaptive Pseudo Augmentation for GAN Training with Limited Data. NeurIPS 2021, 6-14 December 2021, 256-263.
[23] Graves, A., Fernández, S. and Schmidhuber, J. (2007) Multi-Dimensional Recurrent Neural Networks. Artificial Neural Networks-ICANN 2007: 17th International Conference, Porto, 9-13 September 2007, 549-558.

为你推荐