This repository has been archived on 2025-08-26. You can view files and clone it, but cannot push or open issues or pull requests.
Files
enginex_bi_series-sherpa-onnx/sherpa-onnx/csrc/transducer-keyword-decoder.cc
2024-06-19 20:51:57 +08:00

186 lines
6.3 KiB
C++

// sherpa-onnx/csrc/transducer-keywords-decoder.cc
//
// Copyright (c) 2023-2024 Xiaomi Corporation
#include "sherpa-onnx/csrc/transducer-keyword-decoder.h"
#include <algorithm>
#include <cmath>
#include <cstring>
#include <utility>
#include <vector>
#include "sherpa-onnx/csrc/log.h"
#include "sherpa-onnx/csrc/onnx-utils.h"
namespace sherpa_onnx {
TransducerKeywordResult TransducerKeywordDecoder::GetEmptyResult() const {
int32_t context_size = model_->ContextSize();
int32_t blank_id = 0; // always 0
TransducerKeywordResult r;
std::vector<int64_t> blanks(context_size, -1);
blanks.back() = blank_id;
Hypotheses blank_hyp({{blanks, 0}});
r.hyps = std::move(blank_hyp);
return r;
}
void TransducerKeywordDecoder::Decode(
Ort::Value encoder_out, OnlineStream **ss,
std::vector<TransducerKeywordResult> *result) {
std::vector<int64_t> encoder_out_shape =
encoder_out.GetTensorTypeAndShapeInfo().GetShape();
if (encoder_out_shape[0] != result->size()) {
SHERPA_ONNX_LOGE(
"Size mismatch! encoder_out.size(0) %d, result.size(0): %d\n",
static_cast<int32_t>(encoder_out_shape[0]),
static_cast<int32_t>(result->size()));
exit(-1);
}
int32_t batch_size = static_cast<int32_t>(encoder_out_shape[0]);
int32_t num_frames = static_cast<int32_t>(encoder_out_shape[1]);
int32_t vocab_size = model_->VocabSize();
int32_t context_size = model_->ContextSize();
std::vector<int64_t> blanks(context_size, -1);
blanks.back() = 0; // blank_id is hardcoded to 0
std::vector<Hypotheses> cur;
for (auto &r : *result) {
cur.push_back(std::move(r.hyps));
}
std::vector<Hypothesis> prev;
for (int32_t t = 0; t != num_frames; ++t) {
// Due to merging paths with identical token sequences,
// not all utterances have "num_active_paths" paths.
auto hyps_row_splits = GetHypsRowSplits(cur);
int32_t num_hyps =
hyps_row_splits.back(); // total num hyps for all utterance
prev.clear();
for (auto &hyps : cur) {
for (auto &h : hyps) {
prev.push_back(std::move(h.second));
}
}
cur.clear();
cur.reserve(batch_size);
Ort::Value decoder_input = model_->BuildDecoderInput(prev);
Ort::Value decoder_out = model_->RunDecoder(std::move(decoder_input));
Ort::Value cur_encoder_out =
GetEncoderOutFrame(model_->Allocator(), &encoder_out, t);
cur_encoder_out =
Repeat(model_->Allocator(), &cur_encoder_out, hyps_row_splits);
Ort::Value logit =
model_->RunJoiner(std::move(cur_encoder_out), View(&decoder_out));
float *p_logit = logit.GetTensorMutableData<float>();
LogSoftmax(p_logit, vocab_size, num_hyps);
// The acoustic logprobs for current frame
std::vector<float> logprobs(vocab_size * num_hyps);
std::memcpy(logprobs.data(), p_logit,
sizeof(float) * vocab_size * num_hyps);
// now p_logit contains log_softmax output, we rename it to p_logprob
// to match what it actually contains
float *p_logprob = p_logit;
// add log_prob of each hypothesis to p_logprob before taking top_k
for (int32_t i = 0; i != num_hyps; ++i) {
float log_prob = prev[i].log_prob;
for (int32_t k = 0; k != vocab_size; ++k, ++p_logprob) {
*p_logprob += log_prob;
}
}
p_logprob = p_logit; // we changed p_logprob in the above for loop
for (int32_t b = 0; b != batch_size; ++b) {
int32_t frame_offset = (*result)[b].frame_offset;
int32_t start = hyps_row_splits[b];
int32_t end = hyps_row_splits[b + 1];
auto topk =
TopkIndex(p_logprob, vocab_size * (end - start), max_active_paths_);
Hypotheses hyps;
for (auto k : topk) {
int32_t hyp_index = k / vocab_size + start;
int32_t new_token = k % vocab_size;
Hypothesis new_hyp = prev[hyp_index];
float context_score = 0;
auto context_state = new_hyp.context_state;
// blank is hardcoded to 0
// also, it treats unk as blank
if (new_token != 0 && new_token != unk_id_) {
new_hyp.ys.push_back(new_token);
new_hyp.timestamps.push_back(t + frame_offset);
new_hyp.ys_probs.push_back(
exp(logprobs[hyp_index * vocab_size + new_token]));
new_hyp.num_trailing_blanks = 0;
auto context_res = ss[b]->GetContextGraph()->ForwardOneStep(
context_state, new_token);
context_score = std::get<0>(context_res);
new_hyp.context_state = std::get<1>(context_res);
// Start matching from the start state, forget the decoder history.
if (new_hyp.context_state->token == -1) {
new_hyp.ys = blanks;
new_hyp.timestamps.clear();
new_hyp.ys_probs.clear();
}
} else {
++new_hyp.num_trailing_blanks;
}
new_hyp.log_prob = p_logprob[k] + context_score;
hyps.Add(std::move(new_hyp));
} // for (auto k : topk)
auto best_hyp = hyps.GetMostProbable(false);
auto status = ss[b]->GetContextGraph()->IsMatched(best_hyp.context_state);
bool matched = std::get<0>(status);
const ContextState *matched_state = std::get<1>(status);
if (matched) {
float ys_prob = 0.0;
for (int32_t i = 0; i < matched_state->level; ++i) {
ys_prob += best_hyp.ys_probs[i];
}
ys_prob /= matched_state->level;
if (best_hyp.num_trailing_blanks > num_trailing_blanks_ &&
ys_prob >= matched_state->ac_threshold) {
auto &r = (*result)[b];
r.tokens = {best_hyp.ys.end() - matched_state->level,
best_hyp.ys.end()};
r.timestamps = {best_hyp.timestamps.end() - matched_state->level,
best_hyp.timestamps.end()};
r.keyword = matched_state->phrase;
hyps = Hypotheses({{blanks, 0, ss[b]->GetContextGraph()->Root()}});
}
}
cur.push_back(std::move(hyps));
p_logprob += (end - start) * vocab_size;
} // for (int32_t b = 0; b != batch_size; ++b)
}
for (int32_t b = 0; b != batch_size; ++b) {
auto &hyps = cur[b];
auto best_hyp = hyps.GetMostProbable(false);
auto &r = (*result)[b];
r.hyps = std::move(hyps);
r.num_trailing_blanks = best_hyp.num_trailing_blanks;
r.frame_offset += num_frames;
}
}
} // namespace sherpa_onnx