Closed jtdutta1 closed 4 years ago
Hi Jeet!
Can you let me know the dimensions of logits, labels, label_length and logit_length which you have passed in rnnt_loss call?
Hi, All the info is actually given in the log I provided. Sorry I didn't label it properly, but here it goes.:- logits: (4, 391, 172, 28) labels: (4, 172) label_length: (4, 1) logit_length: (4, 1)
Thanks for the fast reply.
I see two issues here.
Firstly, reshape _labellength and _logitlength to (4). Secondly, if your labels have a maximum sequence length (U) of 172. Then, logits should be of the shape (4, 391, 173, 28). Logits should have 3rd dimension as U+1 since then the prediction network operates of [0] + [label_sequence_ids], i.e., with a blank symbol (0) prepended to the actual sequence.
Also, note that ids in labels should be in [1, 28] and not [0, 27]. Since loss_function assumes that index 0 is reserved for the blank symbol. Plus, labels and _labellength exclude prepended blank symbol so their contents should correspond to length U only.
I'll update these in the README.
Thank You! I'll include these changes and will let you know
I got this error.