Supervit for deep learning

This article This is an article published this year Transformer Class article . The author puts forward ViT Enhanced Edition ——SuperViT, Its core is 2 individual ：① Multiscale input ;②Token-keeping-rate Mechanism . The author proved through a series of experiments SuperViT It can achieve the ratio of calculation efficiency and accuracy Swin-T Better expression ！

Reference documents ：
① Source code
②Transformer Study ( Four )—DeiT
③ Various types Transformer All have to be a little inferior ,LV-ViT： Explore multiple for ascension ViT High efficiency of performance Trick