RuntimeException: inferenceWithScores not supported for ViterbiLayer about processors HOT 7 OPEN

I'm finally looking. Sorry for the delay.

from processors.

I wonder whether the scores are really necessary given that access isn't provided. So the hierarchy is

FinalLayer
  <- ForwardLayer
    <- GreedyForwardLayer
    <- ViterbiForwardLayer

and Greedy has it while Viterbi doesn't. It seems like there needs to be a Utils.viterbiWithScores for ViterbiForwardLayer.inferenceWithScores in the same way inference calls Utils.viterbi. That could probably be made by keeping track of backScores along with backPointers.

That probably doesn't answer your question. If you replace inferenceWithScore with just inference won't you lose the scores you are trying to locate? Is it loss() that you need instead?

from processors.

This one is on me. I implemented it for greedy inference, and did not implement it yet for Viterbi.
Haris, for now please change the inference type to greedy in the final layer, in the config file. It should work afterwards.

from processors.

I guess what I really need to return from the viterbi() method in Utils.scala is pathScore along with bestPath. That would involve minor modifications to viterbi() and inference() but this is a short term solution that won't return an IndexedSeq[IndexedSeq[(String, Float)] nested array as specified in the inferenceWithScores() method.

from processors.

This one is on me. I implemented it for greedy inference, and did not implement it yet for Viterbi. Haris, for now please change the inference type to greedy in the final layer, in the config file. It should work afterwards.

ok.

from processors.

@harisriaz17, keep in mind there is one viterbi in org.clulab.dynet.Utils. Perhaps in the end one version can be used for both the existing and the new case.

from processors.

Will this for sure be needed eventually, or is it contingent on things that might not happen? I.e., will implementing it now be potentially wasted effort?

from processors.