Multi-discriminator error back propagation confrontation network method

Abstract

The invention discloses a multi-discriminator error back propagation confrontation network method, and belongs to the depth learning nerve network field; a model is built by the following steps: S1, constructing a formation confrontation network GAN model; S2, constructing a plurality of discriminators on the existing GAN model basis; S3, preparing a data set so as to train an improved multi-discriminator network; S4, recording all discriminator loss functions in each training process; S5, calculating the average value of all discriminator loss functions, and returning to a generator so as tocarry out follow up network training. A single discriminator may cause the network poor robustness problems in the network training process; the method can solve said problems, and can form the multi-discriminator formation confrontation network, thus evaluating the true and false of generator formed images from a more objective angle, and providing better training effect of the whole formation confrontation network.

Description

technical field [0001] The invention relates to the technical field of deep learning neural networks, in particular to an adversarial network method for error backpropagation of multiple discriminators. Background technique [0002] Generative Adversarial Network (GAN for short) is a framework proposed by Goodfellow in 2014. It is based on the idea of ​​"game theory" and constructs two models of generator (generator) and discriminator (discriminator). Uniform noise of (0, 1) or Gaussian random noise generates images, and the latter discriminates the input image to determine whether it is an image from the dataset or an image produced by the generator. Every time the discriminator completes a judgment, it returns the result error to the generator. [0003] However, in the original GAN ​​model, there is only one discriminator, which means that the decision whether the generator generates an image is true or not depends only on this one discriminator. The result of this situa...

AI Technical Summary

Problems solved by technology

The result in this case is that if the discriminator's judgment on the image generated by the generator is biased, it will affect the accuracy and speed of the entire network training

Images