【摘要】 针对生成对抗网络（GAN）训练中的不稳定性问题，提出一种基于梯度范数自适应正则化的一致性优化方法。该方法通过引入依赖梯度范数的正则化强度机制，自适应调控生成器与判别器的优化过程，从而维持双方的训练平衡。该方法对应于一个改进的向量场，并通过分析对应的不动点迭代，从理论上证明了算法的局部收敛性。为验证其有效性，在高斯混合分布数据集以及CIFAR-10、CelebA标准图像数据集上进行了系统实验。结果表明，该方法能够稳定生成高质量、多样化的图像样本，在定量评价指标上优于对比方法。更多还原

【Abstract】 To address the instability issue in the training of Generative Adversarial Network(GAN),a consistency optimization method based on adaptive regularization with gradient norm,termed A-ConOpt,is proposed. This method introduces a regularization strength mechanism dependent on the gradient norm to adaptively regulate the optimization processes of the generator and the discriminator,thereby maintaining the training balance between the two networks. The method corresponds to a modified vector field,and the local convergence of the algorithm is theoretically established through the analysis of its corresponding fixed-point iteration. To verify its effectiveness,systematic experiments are conducted on a Gaussian mixture distribution dataset as well as the standard image datasets CIFAR-10 and CelebA. The results demonstrate that the proposed method can stably generate high-quality and diverse image samples and outperforms the compared methods in terms of quantitative evaluation metrics.更多还原