Closed Hansyvea closed 3 years ago
Hi again @Hansyvea .
Thanks for input! that is also my experience.
I have not experience with adding CRF layers. Do you have a suggestion? With NERDA you can provide your own network, so I think, it could be implemented there?
by observing the prediction output, i found many errors were made by wrong prefix. it is possible that a CRF layer as decoder could solve the problem