Training Generative Adversarial Networks with Adaptive Composite Gradient
This addresses the problem of inefficient and unstable GAN training for researchers and practitioners, offering an incremental improvement over existing methods.
The paper tackles the challenge of training Generative Adversarial Networks (GANs) by proposing the adaptive Composite Gradients (ACG) method, which reduces cyclic behaviors and computational costs, and shows competitive or superior performance on datasets like MNIST, CIFAR-10, and CelebA compared to baselines.
The wide applications of Generative adversarial networks benefit from the successful training methods, guaranteeing that an object function converges to the local minima. Nevertheless, designing an efficient and competitive training method is still a challenging task due to the cyclic behaviors of some gradient-based ways and the expensive computational cost of these methods based on the Hessian matrix. This paper proposed the adaptive Composite Gradients (ACG) method, linearly convergent in bilinear games under suitable settings. Theory and toy-function experiments suggest that our approach can alleviate the cyclic behaviors and converge faster than recently proposed algorithms. Significantly, the ACG method is not only used to find stable fixed points in bilinear games as well as in general games. The ACG method is a novel semi-gradient-free algorithm since it does not need to calculate the gradient of each step, reducing the computational cost of gradient and Hessian by utilizing the predictive information in future iterations. We conducted two mixture of Gaussians experiments by integrating ACG to existing algorithms with Linear GANs. Results show ACG is competitive with the previous algorithms. Realistic experiments on four prevalent data sets (MNIST, Fashion-MNIST, CIFAR-10, and CelebA) with DCGANs show that our ACG method outperforms several baselines, which illustrates the superiority and efficacy of our method.