EngineX-Iluvatar/enginex-mr_series-sherpa-onnx
Archived
16
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Add RNN LM rescore for offline ASR with modified_beam_search (#125)

Browse Source
This commit is contained in:
Fangjun Kuang
2023-04-23 17:15:18 +08:00
committed by GitHub
parent d49a597431
commit 86017f9833
32 changed files with 842 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
sherpa-onnx/csrc/hypothesis.cc
View File

@@ -17,6 +17,9 @@ void Hypotheses::Add(Hypothesis hyp) {
hyps_dict_[key] = std::move(hyp);
} else {
it->second.log_prob = LogAdd<double>()(it->second.log_prob, hyp.log_prob);
it->second.lm_log_prob =
LogAdd<double>()(it->second.lm_log_prob, hyp.lm_log_prob);
}
}
@@ -24,8 +27,8 @@ Hypothesis Hypotheses::GetMostProbable(bool length_norm) const {
if (length_norm == false) {
return std::max_element(hyps_dict_.begin(), hyps_dict_.end(),
[](const auto &left, auto &right) -> bool {
return left.second.log_prob <
right.second.log_prob;
return left.second.TotalLogProb() <
right.second.TotalLogProb();
})
->second;
} else {
@@ -33,8 +36,8 @@ Hypothesis Hypotheses::GetMostProbable(bool length_norm) const {
return std::max_element(
hyps_dict_.begin(), hyps_dict_.end(),
[](const auto &left, const auto &right) -> bool {
return left.second.log_prob / left.second.ys.size() <
right.second.log_prob / right.second.ys.size();
return left.second.TotalLogProb() / left.second.ys.size() <
right.second.TotalLogProb() / right.second.ys.size();
})
->second;
}
@@ -47,15 +50,16 @@ std::vector<Hypothesis> Hypotheses::GetTopK(int32_t k, bool length_norm) const {
std::vector<Hypothesis> all_hyps = Vec();
if (length_norm == false) {
std::partial_sort(
all_hyps.begin(), all_hyps.begin() + k, all_hyps.end(),
[](const auto &a, const auto &b) { return a.log_prob > b.log_prob; });
std::partial_sort(all_hyps.begin(), all_hyps.begin() + k, all_hyps.end(),
[](const auto &a, const auto &b) {
return a.TotalLogProb() > b.TotalLogProb();
});
} else {
// for length_norm is true
std::partial_sort(all_hyps.begin(), all_hyps.begin() + k, all_hyps.end(),
[](const auto &a, const auto &b) {
return a.log_prob / a.ys.size() >
b.log_prob / b.ys.size();
return a.TotalLogProb() / a.ys.size() >
b.TotalLogProb() / b.ys.size();
});
}