zx3Leonardo

hi, I've got the same problem. Why would the author project the features of images before feeding them into attention layer?

I think maybe you need use `imdb = datasets.attr_bpad('test')` to replace original code.