test issue

[vincentwei0919]

hey, I appreciate your nice work, it helps me a lot!
however, I use your code to train my own images which contains 137 classes. After changing your code, labels are correctly made and training seems smoothly. I get some results below and don't know if it works well. And I don't have a monitor, you know, some works are done on servers.
So what can I do to use the results below?

[suraj-deshmukh]

you have a huge number of classes so don't rely on fully correct output. and you will get fully correct output when predicted sequence exactly matches with the original sequence like in my case for a particular image if the original target sequence is [0 1 0 1 0] and predicted sequence is [0 1 0 1 0] then its an exact match. you will need to use a hamming loss to check the accuracy and you also need to try different model architecture to get better results.