Hi, I’m trying to use .generate()
method for decoder-only models(ex. GPT2), and figuring out that it provides various helpful values as well as the generated sequence(s). However, I cannot find detailed description of each value.
In detail, when I use beam search (by passing num_beams
3, num_return-sequences
3, do_sample
False), I can get sequences
, sequences_scores
, scores
, and beam_indices
. Obviously, sequences
is the generated sequences(token ids). But what do others mean? And how can I compute them by hand?
Plus, huggingface transformers provide compute_transition_scores()
method for each model. What does this value mean? How can I compute it? And what’s the difference between scores
and transition_scores
?