Semi-supervised learning method for GAN model training

Abstract

The invention discloses a semi-supervised learning method for GAN model training through adaptive supervised ratio control. According to the method, the problem that the authenticity identification rule obtained by adversarial device learning in the GAN model training process is insufficient to guide a generator to generate a high-precision case is effectively solved; through the self-adaptive supervision ratio, the model automatically controls the amount of fused label information, the data information is effectively utilized, and the generated case meets the diversity and fidelity at the same time.

Description

technical field [0001] The invention relates to a GAN model training method in the field of deep learning, in particular to a semi-supervised GAN learning method for controlling the degree of supervision through an adaptive supervision ratio. Background technique [0002] Generative Adversarial Networks (GAN, Generative Adversarial Networks) is a deep learning model. GAN models are widely used in image generation, speech generation, text generation and other fields. For the training of GAN models, unsupervised training methods are generally used, so that a large amount of unlabeled natural data can be used; when there is partly labeled data, the semi-supervised learning framework can be used to obtain the input code and label of the generator. Correspondence, complete the training of C-GAN. However, these general methods have the following problems: 1. The data information, especially the label information, cannot be fully utilized; 2. The authenticity rules learned by the ...

Images