A Dynamic Adjustment Method Based on DCGAN Performance

Abstract

The invention discloses a dynamic adjustment method based on DCGAN performance, which belongs to the field of deep learning signal processing. The dynamic adjustment method includes the following steps: S1, constructing a deep convolution generation confrontation network DCGAN; S2, using an image data set to perform network Training; S3, using the network loss function to define the performance of the generator and the discriminator; S4, defining the ability of the network to generate images according to the performance of the generator and the discriminator; S5, monitoring the network training in real time to realize the dynamic adjustment of the network performance. This method can monitor the performance of DCGAN in real time, realize the dynamic adjustment of the network's ability to generate images, ensure the balanced improvement of the generator and discriminator capabilities, and avoid the "mode collapse" in the network training process, so that The quality of images generated by DCGAN is guaranteed.

Description

technical field [0001] The invention relates to the technical field of deep learning signal processing, in particular to a dynamic adjustment method based on DCGAN performance. 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. Using the return error of the discriminator, the generator further improves its own model to generate higher-quality images. When the generator cannot distinguish whether the image comes from the data set or the generator, it is considered that the generato...

AI Technical Summary

Problems solved by technology

However, it is worth noting that there is no unified measure of the "ability" of the generator and discriminator, which means that the training of the entire network can continue indefinitely

In the initial stage of training, the abilities of both parties are constantly improving, but as the number of training increases, the overlap rate of the input samples of the generator will continue to increase, and the gradient value will become smaller and smaller, that is, it is difficult to further improve the ability. Limited, the quality of the image input to the discriminator has not been improved, and the ability to improve the discriminator is also limited

With the continuous training of the network, the generator will also appear to be overfitting, that is, when the text in the data set is input, the quality of the generated image is very good, but once the user is allowed to input the text description, the quality of the generated image is difficult to obtain ensure

Although the loss function is defined, there is no quantitative evaluation standard based on parameters such as the loss function and the number of training times, and the loss functions of both parties have not been unified in the overall framework, that is, there is no ability to combine the generator and the discriminator. Linked together to describe, that is, the measurement standard of the entire network framework is missing

The result of this state of lack of an evaluation standard mechanism is that the entire network is endlessly trained

More importantly, there is no balanced criterion for the capabilities of the generator and the discriminator, so that during the training process, the ability of one will far exceed the other, resulting in the collapse of the entire network model

Images