att, original_att and w2v_att problems

[zjrzjr666]

hi, it was a pleasure to read your paper. I had some problems reading your code. For example, in the first figure, I noticed that you seem to have extracted three attention vectors, att, original_att and w2v_att. The third one I can understand is that you get the class attribute through w2v? But your paper seems to use GloVe. And can you tell me the difference and meaning of att and original_att? They both seem to be (312,200）. I also found class description files in the cub public dataset What is their difference?And why would you read three attribute-based attentions?

[FrankLeeCode]

Here's the explanation in readme file of xlsa17. Hope it can help.

att_splits.mat includes the following fields:

• att: columns correpond to class attribute vectors normalized to have unit l2 norm, following the classes order in allclasses.txt
• original_att: the original class attribute vectors without normalization
• trainval_loc: instances indexes of train+val set features (for only seen classes) in resNet101.mat
• test_seen_loc: instances indexes of test set features for seen classes
• test_unseen_loc: instances indexes of test set features for unseen classes

[hbdat]

Hi @FrankLeeCode ,

Thank you so the answer :)

Best, Dat