Open sangrockEG opened 4 years ago
I've got the same problem. And I extracted train_loc and val_loc from trainval_loc according to "trainclasses.txt" and "valclasses.txt". But the result is not so good, especially at test_seen.
I did not test with train/val split of trainval class, so thanks for your comment. Is your result much worse than the result of paper? It's really weird.. may only the authors can help us.
The 'gzsl unseen' is 34. while it is 39.3 in the paper. And the 'gzsl seen' is only 26. while it is 66.7 in the paper. The classifier is knn so I even try to train knn with trainval_loc and test with test_seen, but it only got 47. I'm very confused...
The dataset CADA-VAE provides works well.
I downloaded xlsa17 proposed split dataset to test CUB. But att_mat does not contain neither train_loc or val_loc keys. It only includes trainval_loc. Funny thing is, in standard split, there exists train_loc and val_loc. Is it intended?