I’d like to compute the sequences_scores for arbitrary sequences that aren’t necessarily beam search outputs so I want to understand what this computation is doing. Looking at the numbers it seems like the sequences_scores are different from the losses.
Hi @brianbargh 
The scores and the losses are different things – the former is akin to probabilities, while the latter depends on which loss you’ve picked for your model.
See this thread for the discussion about the computation of sequence_scores 