Generative adversarial network training method and device

Abstract

The embodiment of the invention provides a generative adversarial network training method and device, and the method comprises the steps: fixing parameters of a discriminator model, and carrying out iterative optimization of the parameters of a generator model through a loss function till the similarity of the generator model reaches a first threshold value; fixing the parameters of the generatormodel, and iteratively optimizing the parameters of the discriminator model through a loss function until the discrimination rate of the discriminator model reaches a second threshold; wherein the value of the first threshold value and the value of the second threshold value are both related to the current alternating frequency; adding 1 to the alternating times; repeating the above steps until the discriminator model and the generator model reach Nash equilibrium; and determining the trained generative adversarial network according to the current discrimination model and the current generation model. According to the embodiment of the invention, the iteration time of the generator model and the discriminator model can be controlled, the training efficiency of the generative adversarial model can be improved, and the situation of model crash is avoided.

Description

technical field [0001] Embodiments of the present invention relate to the technical field of artificial intelligence, and in particular to a training method and equipment for generating an adversarial network. Background technique [0002] Generative Adversarial Networks (GAN) is a recently developed unsupervised deep learning model. The core idea of ​​this method is a two-player game (Two-player Game), and its two players are composed of a generative model (Gennerative Model) and a discriminative model (Discriminative Model). [0003] In the prior art, the GAN model is usually trained with a fixed number of iterations. [0004] However, adopting a fixed number of training times will lead to a very time-consuming training process and prone to model collapse. Contents of the invention [0005] Embodiments of the present invention provide a training method and device for generating an adversarial network, so as to improve training efficiency and avoid model collapse. [0...

AI Technical Summary

Problems solved by technology

[0004] However, using a fixed number of training times will lead to a very time-consuming training process and prone to model crashes

Images