Performance Analysis of Different GAN Models: DC-GAN and LS-GAN

Generative Adversarial Networks (GAN) is an excellent method to generate fake images using the training dataset provided. Generative Adversarial Network consists of two sub-networks that is generator and discriminator networks that compete against each to produce the desired output. Regular GAN is not compatible enough for synthesizing images. However, with the use of CNN models for the sub-networks, it is easier to learn the distribution of training images. We demonstrate the performance of CNN based DC-GAN model. However, it can be noticed that sigmoid based cross entropy loss function results in mode collapse and convergence problems. Hence, we further investigated the research using LS-GAN model which uses least square error loss function. Both the models were trained on the same training datasets. To validate the results, generator & discriminator loss and FID scores were calculated. In this paper, fthe different datasets are used to carry out the experiments. Comparison between DC-GAN and LS-GAN models, the impact of changing hyperparameters and the significance of calculated metrics is illustrated.