Analysis of Hierarchical Generative Adversarial Network Model Based on Shifted Window
This paper describes a hierarchical generative adversarial network model based on a sliding window.To address the issues of existing transformer-based generative adversarial network models,such as the inability to obtain multi-scale features,the model is improved by combining sliding windows and window masking strategies.By using sliding windows and window partitioning in the generator and discriminator,the computational complexity of attention calculation is significantly reduced,while achieving mutual communication between attention windows.Furthermore,by adding a sliding window attention mask to the attention module,batch attention calculation for attention windows is achieved while ensuring that the semantic information of the overall feature map is not destroyed.The effectiveness of this method is demonstrated through experiments on multiple datasets in the article.
NETL
NSTL
万方数据
