Exploring the Effectiveness of Hyperparameters in Deep Convolution Generative Adversarial Networks

Traditional machine learning models are often only able to classify or predict data, and cannot generate new simulated data, while Generative Adversarial Network (GAN) provides the possibility for machines to generate high-quality and diverse data. GAN can be used for image data generation, image-to-image transformation, translation of image information and text information into each other, and so on. There are many different types of GAN can realize different types of functions. GAN has a strong generating ability, can generate some high-quality simulation data for human reference. The framework idea of GAN is also very innovative and has been widely developed for applications in unsupervised or semi-supervised learning. In addition, GAN can be integrated with other machine learning models to form a more powerful model, so as to solve the problems that are difficult to deal with traditional models. Therefore, GAN is a very promising model in the field of artificial intelligence. In this thesis, the author completed the basic reproduction of Deep Convolution Generative Adversarial Network (DCGAN), and analyzed the quality of DCGAN generated images, using the control variable method. Some parameters are modified, such as learning rate, batch size, latent size, and explored the influence of these parameters on the training effect of DCGAN. The characteristics of basic parameters on DCGAN image generation results are summarized.