Training method of image generator and image generation method and device

Abstract

The embodiment of the invention discloses a training method of an image generator, and an image generation method and device. The training method of the image generator comprises the steps of generating an initial generative adversarial network model based on sparseness of the generator and a discriminator; and performing iterative training on the initial generative adversarial network model based on the sample data, updating a connection relation of each network layer in the current iterative generator in the iterative training process, continuing to perform iterative training on the updated generator and the current iterative discriminator until an end condition is met, and determining the trained generator as a target image generator. In the training process, connection with low importance is updated, so that pruning on the connection is realized, and the image generator considering processing precision and sparseness is obtained after training is finished. Meanwhile, the sparseness of the generator is kept unchanged, so that the calculated amount in the training process is reduced, and the training efficiency is improved due to few training parameters.

Description

technical field [0001] Embodiments of the present invention relate to the technical field of deep learning, and in particular, to a training method for an image generator, an image generation method and an apparatus. Background technique [0002] Generative Adversarial Networks (GAN, Generative Adversarial Networks) is a deep learning model and one of the most promising methods for unsupervised learning on complex distributions in recent years. The model produces quite good output through the mutual game learning of (at least) two modules in the framework: the generator (Generative Model) and the discriminator (Discriminative Model). [0003] As the quality of image generation improves, so does the training cost of GANs. At present, the compressed generative adversarial network can be obtained through the model compression technology of the generative adversarial network. Among them, the model compression technology includes pruning technology, distillation technology, lott...

AI Technical Summary

Problems solved by technology

[0004] In the process of realizing the present invention, it is found that there are at least the following technical problems in the prior art: a dense generation confrontation network needs to be trained before model compression, and the process of model compression includes the process of compression and retraining performed iteratively, resulting in The overall calculation amount far exceeds the training of the dense model

Images