raidasgrisk / tf2-crnn Goto Github PK

I find you make blank_index equals the 62(last char), but when you decode, you ignore the 0, so the net can't predict the label "0".
I think you should change the char_dict, place the blank label in the first, and make blank_index equals 0, it will work well.

By the way, the output of the network is:
[62 62 62 62 62 62 62 62 62 62 48 17 17 17 17 18 18 18 18 62 62 62 62 62
62 62 62 62 62 62 62 62 62 21 21 22 62 62 62 62 62 62 62 62 62 62 62]
because the zero-padding the output of ctc_greedy_decoder will be:
[48 17 18 21 22 0 0 0 0 0]

hello, when I train my own dataset, the training loss is always Nan, how to resolve this problem?