It is trivial to calculate the CTC loss of a sequence with all blanks. But tf.nn.ctc_loss cannot handle the situation that one or more sequences in the batch have no non-blank labels. This is a big limitation.

P.S.: I pulled these requests before:
I reported this previously at
#13457
I asked this on stackoverflow
https://stackoverflow.com/questions/46652720/how-to-calculate-ctc-loss-of-a-sequence-with-all-blanks-using-tf-nn-ctc-loss