whisper : improve beam search candidate diversity (#1947)

As of #1486, whisper.cpp uses a unified KV cache with KQ masking.
As a result, depending on their location in the batch,
identical sequences in a batch can have slightly different outputs
due to floating point rounding errors during reduction.
See the discussion in #1941 for more details.

The beam search code used "has identical sum of log probabilities"
as a shorthand for "is an identical token sequence". However, per above,
identical tokens do not necessarily result in identical probabilities.

Instead, explicitly compare on sequences.
This is linear in cost when they are identical,
but the lengths are always small and the comparisons are cheap.

This increases diversity during beam search.

This improves output quality for some short samples I've been working
with, at no detectable performance cost.
I haven't checked against larger corpuses.

Fixes #1941
pull/1913/head^2
Josh Bleecher Snyder 2024-03-10 07:54:43 -07:00 committed by GitHub
parent ccf022f970
commit ec166499d8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 14 additions and 1 deletions

View File

@ -4759,6 +4759,19 @@ static void whisper_process_logits(
#endif
}
static bool whisper_sequence_tokens_equal(const whisper_sequence & a, const whisper_sequence & b) {
if (a.tokens.size() != b.tokens.size()) {
return false;
}
// sequences are more likely to diverge at the end
for (int i = a.tokens.size() - 1; i >= 0; i--) {
if (a.tokens[i].id != b.tokens[i].id) {
return false;
}
}
return true;
}
static whisper_token_data whisper_sample_token(
whisper_context & ctx,
const whisper_decoder & decoder,
@ -5378,7 +5391,7 @@ int whisper_full_with_state(
auto & cur = beam_candidates[cur_c++];
while (beam_candidates.size() > cur_c && beam_candidates[cur_c].sequence.sum_logprobs_all == cur.sequence.sum_logprobs_all && i > 0) {
while (beam_candidates.size() > cur_c && whisper_sequence_tokens_equal(beam_candidates[cur_c].sequence, cur.sequence) && i > 0) {
++cur_c;
}