xuning / sherpaonnx

转到一个项目
Toggle navigation Toggle navigation pinning
Fangjun Kuang
Committed by GitHub
aa48b76d 1 parent aeb112dd

Fix initial tokens to decoding (#246)

内嵌 并排对比
正在显示 5 个修改的文件 包含 10 行增加 和 5 行删除
1 cmake_minimum_required(VERSION 3.13 FATAL_ERROR) 1 cmake_minimum_required(VERSION 3.13 FATAL_ERROR)
2 project(sherpa-onnx) 2 project(sherpa-onnx)
3 3
4 -set(SHERPA_ONNX_VERSION "1.6.1") 4 +set(SHERPA_ONNX_VERSION "1.6.2")
5 5
6 # Disable warning about 6 # Disable warning about
7 # 7 #
@@ -30,8 +30,9 @@ OfflineTransducerGreedySearchDecoder::Decode(Ort::Value encoder_out, @@ -30,8 +30,9 @@ OfflineTransducerGreedySearchDecoder::Decode(Ort::Value encoder_out,
30 30
31 std::vector<OfflineTransducerDecoderResult> ans(batch_size); 31 std::vector<OfflineTransducerDecoderResult> ans(batch_size);
32 for (auto &r : ans) { 32 for (auto &r : ans) {
  33 + r.tokens.resize(context_size, -1);
33 // 0 is the ID of the blank token 34 // 0 is the ID of the blank token
34 - r.tokens.resize(context_size, 0); 35 + r.tokens.back() = 0;
35 } 36 }
36 37
37 auto decoder_input = model_->BuildDecoderInput(ans, ans.size()); 38 auto decoder_input = model_->BuildDecoderInput(ans, ans.size());
@@ -32,7 +32,8 @@ OfflineTransducerModifiedBeamSearchDecoder::Decode( @@ -32,7 +32,8 @@ OfflineTransducerModifiedBeamSearchDecoder::Decode(
32 int32_t vocab_size = model_->VocabSize(); 32 int32_t vocab_size = model_->VocabSize();
33 int32_t context_size = model_->ContextSize(); 33 int32_t context_size = model_->ContextSize();
34 34
35 - std::vector<int64_t> blanks(context_size, 0); 35 + std::vector<int64_t> blanks(context_size, -1);
  36 + blanks.back() = 0;
36 37
37 std::deque<Hypotheses> finalized; 38 std::deque<Hypotheses> finalized;
38 std::vector<Hypotheses> cur; 39 std::vector<Hypotheses> cur;
@@ -55,7 +55,8 @@ OnlineTransducerGreedySearchDecoder::GetEmptyResult() const { @@ -55,7 +55,8 @@ OnlineTransducerGreedySearchDecoder::GetEmptyResult() const {
55 int32_t context_size = model_->ContextSize(); 55 int32_t context_size = model_->ContextSize();
56 int32_t blank_id = 0; // always 0 56 int32_t blank_id = 0; // always 0
57 OnlineTransducerDecoderResult r; 57 OnlineTransducerDecoderResult r;
58 - r.tokens.resize(context_size, blank_id); 58 + r.tokens.resize(context_size, -1);
  59 + r.tokens.back() = blank_id;
59 60
60 return r; 61 return r;
61 } 62 }
@@ -42,7 +42,9 @@ OnlineTransducerModifiedBeamSearchDecoder::GetEmptyResult() const { @@ -42,7 +42,9 @@ OnlineTransducerModifiedBeamSearchDecoder::GetEmptyResult() const {
42 int32_t context_size = model_->ContextSize(); 42 int32_t context_size = model_->ContextSize();
43 int32_t blank_id = 0; // always 0 43 int32_t blank_id = 0; // always 0
44 OnlineTransducerDecoderResult r; 44 OnlineTransducerDecoderResult r;
45 - std::vector<int64_t> blanks(context_size, blank_id); 45 + std::vector<int64_t> blanks(context_size, -1);
  46 + blanks.back() = blank_id;
  47 +
46 Hypotheses blank_hyp({{blanks, 0}}); 48 Hypotheses blank_hyp({{blanks, 0}});
47 r.hyps = std::move(blank_hyp); 49 r.hyps = std::move(blank_hyp);
48 r.tokens = std::move(blanks); 50 r.tokens = std::move(blanks);