Invalid argument: indices[0,0,2] = [0, 0, 2, -1] does not index into shape [10,79,18,6484] about rnnt HOT 2 OPEN

I run your Sample Train Script and found that you set lables=decoder_seqs + 1 in
return tf.reduce_mean(self.loss(log_probs, decoder_seqs + 1, decoder_lens, encoder_lens))
If i change your code to return tf.reduce_mean(self.loss(log_probs, decoder_seqs, decoder_lens, encoder_lens)), it lead to the same error.
I can't understand why you do that.
In my dictionary, I reserve index 0 for blank, and the word indexed from 1 to 6484 (6485 at all), so my decoder_seqs represent the real index of word, why plus 1?

from rnnt.

my labels is padded with 0 as follows:
[[ 379 57 230 88 214 268 907 1061 62 142 2027 0 0 0 0 0] [ 588 459 357 547 209 102 101 86 3 96 87 89 32 0 0 0] [ 144 942 196 571 1956 219 82 908 46 363 1148 696 634 55 761 383] [ 624 316 378 716 1737 362 49 110 746 1157 346 380 570 0 0 0] [ 781 260 1311 1562 15 429 16 666 150 0 0 0 0 0 0 0] [ 809 411 310 178 541 1209 2 99 102 87 232 88 231 0 0 0] [ 704 6 3 570 167 151 3 290 101 102 87 89 2287 0 0 0] [ 3 645 146 3748 4067 4735 869 60 1407 95 1334 4735 869 0 0 0] [ 52 342 372 1020 440 441 628 440 441 628 1154 2121 378 233 766 0] [1180 478 1087 1687 2381 777 174 231 102 97 174 0 0 0 0 0]]

from rnnt.