Open yzhang-github-pub opened 1 year ago
@yzhang-github-pub do you mean beam search?
@lucidrains Not exactly. Using language translation as an example. The model generates one best target sentence, either greedily or by beam search. What I ask is to generate top k target sentences.
For me, it would actually be great if generate
had a beam search option, since I need multiple outputs ranked by probability for character-level generation of phonetic transcriptions.
@maryszmary yea i should add classic beam search at some point for sure
maybe roll up any interesting advancements in that area of research with its release? lemme know
Instead of generate the best (top 1) sequence from autoregressive generate function, it is useful to generate >1 best sequences. For example, in language translation, human users can have choices from more than 1 best translations.
Thanks for your consideration.