Some questions about unimportant token selection

[studentyyh]

Thank you very much for your good work. I have encountered a problem when reading the paper and checking the code, and I would like to ask you about it.
In the paper optimization stage of the insignificant token selection, the paper diagram shows that it is selected from the token after encoder, but why is the code selected from the feature token after patch_embeding? Looking forward to your reply.