System and method for training generative adversarial network and based on reverse teacher, electronic device, storage medium, and program product

Abstract

The present invention discloses a system for training a generative adversarial network model. The system includes an optimal generator module which is used for generating synthetic data that is highly consistent with a real data distribution; a discriminator group module which includes a plurality of discriminators with different structures and different initial parameters, and is used for distinguishing between real data and the synthetic data, and predicting a target value; and an auxiliary scorer module which is used for carrying out quantitative evaluation on performance of generators and the discriminators. According to the system, through dynamically monitoring a performance gap between the optimal generator module and the discriminator group module, a reverse teacher mechanism is triggered when preset conditions are met, so that a student model in the optimal generator module and the discriminator group module teaches a teacher model in the optimal generator module and the discriminator group module, and a dynamic performance balance between the optimal generator module and the discriminator group module is maintained. According to the present invention, problems of easily falling into local optimum, mode collapse, and instable training in a process of training a generative adversarial network are effectively solved.

Images