Support batch greedy search decoding (#30)

Fangjun Kuang · GitHub
Commit 8acc059b3f5d88fc804408ce91b6a32b564d56b1 8acc059b 1 parent ebc3b47f
sherpa-onnx/csrc/online-lstm-transducer-model.cc
sherpa-onnx/csrc/online-lstm-transducer-model.h
sherpa-onnx/csrc/online-recognizer.cc
sherpa-onnx/csrc/online-transducer-greedy-search-decoder.cc
sherpa-onnx/csrc/online-transducer-model.h
--- a/sherpa-onnx/csrc/online-lstm-transducer-model.cc
查看文件 @8acc059
+++ b/sherpa-onnx/csrc/online-lstm-transducer-model.cc
查看文件 @8acc059
@@ -3,6 +3,7 @@
 // Copyright (c)  2023  Xiaomi Corporation
 #include "sherpa-onnx/csrc/online-lstm-transducer-model.h"
+#include <algorithm>
 #include <memory>
 #include <sstream>
 #include <string>
@@ -10,6 +11,7 @@
 #include <vector>
 #include "onnxruntime_cxx_api.h"  // NOLINT
+#include "sherpa-onnx/csrc/online-transducer-decoder.h"
 #include "sherpa-onnx/csrc/onnx-utils.h"
 #define SHERPA_ONNX_READ_META_DATA(dst, src_key)                        \
@@ -114,23 +116,85 @@ void OnlineLstmTransducerModel::InitJoiner(const std::string &filename) {
   }
 }
-Ort::Value OnlineLstmTransducerModel::StackStates(
-    const std::vector<Ort::Value> &states) const {
-  fprintf(stderr, "implement me: %s:%d!\n", __func__,
-          static_cast<int>(__LINE__));
-  auto memory_info =
-      Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-  int64_t a;
-  std::array<int64_t, 3> x_shape{1, 1, 1};
-  Ort::Value x = Ort::Value::CreateTensor(memory_info, &a, 0, &a, 0);
-  return x;
+std::vector<Ort::Value> OnlineLstmTransducerModel::StackStates(
+    const std::vector<std::vector<Ort::Value>> &states) const {
+  int32_t batch_size = static_cast<int32_t>(states.size());
+
+  std::array<int64_t, 3> h_shape{num_encoder_layers_, batch_size, d_model_};
+  Ort::Value h = Ort::Value::CreateTensor<float>(allocator_, h_shape.data(),
+                                                 h_shape.size());
+
+  std::array<int64_t, 3> c_shape{num_encoder_layers_, batch_size,
+                                 rnn_hidden_size_};
+
+  Ort::Value c = Ort::Value::CreateTensor<float>(allocator_, c_shape.data(),
+                                                 c_shape.size());
+
+  float *dst_h = h.GetTensorMutableData<float>();
+  float *dst_c = c.GetTensorMutableData<float>();
+
+  for (int32_t layer = 0; layer != num_encoder_layers_; ++layer) {
+    for (int32_t i = 0; i != batch_size; ++i) {
+      const float *src_h =
+          states[i][0].GetTensorData<float>() + layer * d_model_;
+
+      const float *src_c =
+          states[i][1].GetTensorData<float>() + layer * rnn_hidden_size_;
+
+      std::copy(src_h, src_h + d_model_, dst_h);
+      std::copy(src_c, src_c + rnn_hidden_size_, dst_c);
+
+      dst_h += d_model_;
+      dst_c += rnn_hidden_size_;
+    }
+  }
+
+  std::vector<Ort::Value> ans;
+
+  ans.reserve(2);
+  ans.push_back(std::move(h));
+  ans.push_back(std::move(c));
+
+  return ans;
 }
-std::vector<Ort::Value> OnlineLstmTransducerModel::UnStackStates(
-    Ort::Value states) const {
-  fprintf(stderr, "implement me: %s:%d!\n", __func__,
-          static_cast<int>(__LINE__));
-  return {};
+std::vector<std::vector<Ort::Value>> OnlineLstmTransducerModel::UnStackStates(
+    const std::vector<Ort::Value> &states) const {
+  int32_t batch_size = states[0].GetTensorTypeAndShapeInfo().GetShape()[1];
+
+  std::vector<std::vector<Ort::Value>> ans(batch_size);
+
+  // allocate space
+  std::array<int64_t, 3> h_shape{num_encoder_layers_, 1, d_model_};
+  std::array<int64_t, 3> c_shape{num_encoder_layers_, 1, rnn_hidden_size_};
+
+  for (int32_t i = 0; i != batch_size; ++i) {
+    Ort::Value h = Ort::Value::CreateTensor<float>(allocator_, h_shape.data(),
+                                                   h_shape.size());
+    Ort::Value c = Ort::Value::CreateTensor<float>(allocator_, c_shape.data(),
+                                                   c_shape.size());
+    ans[i].push_back(std::move(h));
+    ans[i].push_back(std::move(c));
+  }
+
+  for (int32_t layer = 0; layer != num_encoder_layers_; ++layer) {
+    for (int32_t i = 0; i != batch_size; ++i) {
+      const float *src_h = states[0].GetTensorData<float>() +
+                           layer * batch_size * d_model_ + i * d_model_;
+      const float *src_c = states[1].GetTensorData<float>() +
+                           layer * batch_size * rnn_hidden_size_ +
+                           i * rnn_hidden_size_;
+
+      float *dst_h = ans[i][0].GetTensorMutableData<float>() + layer * d_model_;
+      float *dst_c =
+          ans[i][1].GetTensorMutableData<float>() + layer * rnn_hidden_size_;
+
+      std::copy(src_h, src_h + d_model_, dst_h);
+      std::copy(src_c, src_c + rnn_hidden_size_, dst_c);
+    }
+  }
+
+  return ans;
 }
 std::vector<Ort::Value> OnlineLstmTransducerModel::GetEncoderInitStates() {
@@ -189,16 +253,21 @@ OnlineLstmTransducerModel::RunEncoder(Ort::Value features,
 }
 Ort::Value OnlineLstmTransducerModel::BuildDecoderInput(
-    const std::vector<int64_t> &hyp) {
-  auto memory_info =
-      Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-  std::array<int64_t, 2> shape{1, context_size_};
+    const std::vector<OnlineTransducerDecoderResult> &results) {
+  int32_t batch_size = static_cast<int32_t>(results.size());
+  std::array<int64_t, 2> shape{batch_size, context_size_};
+  Ort::Value decoder_input =
+      Ort::Value::CreateTensor<int64_t>(allocator_, shape.data(), shape.size());
+  int64_t *p = decoder_input.GetTensorMutableData<int64_t>();
+
+  for (const auto &r : results) {
+    const int64_t *begin = r.tokens.data() + r.tokens.size() - context_size_;
+    const int64_t *end = r.tokens.data() + r.tokens.size();
+    std::copy(begin, end, p);
+    p += context_size_;
+  }
-  return Ort::Value::CreateTensor(
-      memory_info,
-      const_cast<int64_t *>(hyp.data() + hyp.size() - context_size_),
-      context_size_, shape.data(), shape.size());
+  return decoder_input;
 }
 Ort::Value OnlineLstmTransducerModel::RunDecoder(Ort::Value decoder_input) {
--- a/sherpa-onnx/csrc/online-lstm-transducer-model.h
查看文件 @8acc059
+++ b/sherpa-onnx/csrc/online-lstm-transducer-model.h
查看文件 @8acc059
@@ -19,16 +19,19 @@ class OnlineLstmTransducerModel : public OnlineTransducerModel {
  public:
   explicit OnlineLstmTransducerModel(const OnlineTransducerModelConfig &config);
-  Ort::Value StackStates(const std::vector<Ort::Value> &states) const override;
+  std::vector<Ort::Value> StackStates(
+      const std::vector<std::vector<Ort::Value>> &states) const override;
-  std::vector<Ort::Value> UnStackStates(Ort::Value states) const override;
+  std::vector<std::vector<Ort::Value>> UnStackStates(
+      const std::vector<Ort::Value> &states) const override;
   std::vector<Ort::Value> GetEncoderInitStates() override;
   std::pair<Ort::Value, std::vector<Ort::Value>> RunEncoder(
       Ort::Value features, std::vector<Ort::Value> &states) override;
-  Ort::Value BuildDecoderInput(const std::vector<int64_t> &hyp) override;
+  Ort::Value BuildDecoderInput(
+      const std::vector<OnlineTransducerDecoderResult> &results) override;
   Ort::Value RunDecoder(Ort::Value decoder_input) override;
@@ -41,6 +44,7 @@ class OnlineLstmTransducerModel : public OnlineTransducerModel {
   int32_t ChunkShift() const override { return decode_chunk_len_; }
   int32_t VocabSize() const override { return vocab_size_; }
+  OrtAllocator *Allocator() override { return allocator_; }
  private:
   void InitEncoder(const std::string &encoder_filename);
@@ -50,7 +54,6 @@ class OnlineLstmTransducerModel : public OnlineTransducerModel {
  private:
   Ort::Env env_;
   Ort::SessionOptions sess_opts_;
-
   Ort::AllocatorWithDefaultOptions allocator_;
   std::unique_ptr<Ort::Session> encoder_sess_;
--- a/sherpa-onnx/csrc/online-recognizer.cc
查看文件 @8acc059
+++ b/sherpa-onnx/csrc/online-recognizer.cc
查看文件 @8acc059
@@ -6,6 +6,7 @@
 #include <assert.h>
+#include <algorithm>
 #include <memory>
 #include <sstream>
 #include <utility>
@@ -64,39 +65,50 @@ class OnlineRecognizer::Impl {
   }
   void DecodeStreams(OnlineStream **ss, int32_t n) {
-    if (n != 1) {
-      fprintf(stderr, "only n == 1 is implemented\n");
-      exit(-1);
-    }
-    OnlineStream *s = ss[0];
-    assert(IsReady(s));
-
     int32_t chunk_size = model_->ChunkSize();
     int32_t chunk_shift = model_->ChunkShift();
-    int32_t feature_dim = s->FeatureDim();
+    int32_t feature_dim = ss[0]->FeatureDim();
+
+    std::vector<OnlineTransducerDecoderResult> results(n);
+    std::vector<float> features_vec(n * chunk_size * feature_dim);
+    std::vector<std::vector<Ort::Value>> states_vec(n);
-    std::array<int64_t, 3> x_shape{1, chunk_size, feature_dim};
+    for (int32_t i = 0; i != n; ++i) {
+      std::vector<float> features =
+          ss[i]->GetFrames(ss[i]->GetNumProcessedFrames(), chunk_size);
+
+      ss[i]->GetNumProcessedFrames() += chunk_shift;
+
+      std::copy(features.begin(), features.end(),
+                features_vec.data() + i * chunk_size * feature_dim);
+
+      results[i] = std::move(ss[i]->GetResult());
+      states_vec[i] = std::move(ss[i]->GetStates());
+    }
     auto memory_info =
         Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-    std::vector<float> features =
-        s->GetFrames(s->GetNumProcessedFrames(), chunk_size);
-
-    s->GetNumProcessedFrames() += chunk_shift;
+    std::array<int64_t, 3> x_shape{n, chunk_size, feature_dim};
-    Ort::Value x =
-        Ort::Value::CreateTensor(memory_info, features.data(), features.size(),
-                                 x_shape.data(), x_shape.size());
+    Ort::Value x = Ort::Value::CreateTensor(memory_info, features_vec.data(),
+                                            features_vec.size(), x_shape.data(),
+                                            x_shape.size());
-    auto pair = model_->RunEncoder(std::move(x), s->GetStates());
+    auto states = model_->StackStates(states_vec);
-    s->SetStates(std::move(pair.second));
-    std::vector<OnlineTransducerDecoderResult> results = {s->GetResult()};
+    auto pair = model_->RunEncoder(std::move(x), states);
     decoder_->Decode(std::move(pair.first), &results);
-    s->SetResult(results[0]);
+
+    std::vector<std::vector<Ort::Value>> next_states =
+        model_->UnStackStates(pair.second);
+
+    for (int32_t i = 0; i != n; ++i) {
+      ss[i]->SetResult(results[i]);
+      ss[i]->SetStates(std::move(next_states[i]));
+    }
   }
   OnlineRecognizerResult GetResult(OnlineStream *s) {
--- a/sherpa-onnx/csrc/online-transducer-greedy-search-decoder.cc
查看文件 @8acc059
+++ b/sherpa-onnx/csrc/online-transducer-greedy-search-decoder.cc
查看文件 @8acc059
@@ -32,6 +32,30 @@ static Ort::Value GetFrame(Ort::Value *encoder_out, int32_t t) {
       encoder_out_dim, shape.data(), shape.size());
 }
+static Ort::Value Repeat(OrtAllocator *allocator, Ort::Value *cur_encoder_out,
+                         int32_t n) {
+  if (n == 1) {
+    return std::move(*cur_encoder_out);
+  }
+
+  std::vector<int64_t> cur_encoder_out_shape =
+      cur_encoder_out->GetTensorTypeAndShapeInfo().GetShape();
+
+  std::array<int64_t, 2> ans_shape{n, cur_encoder_out_shape[1]};
+
+  Ort::Value ans = Ort::Value::CreateTensor<float>(allocator, ans_shape.data(),
+                                                   ans_shape.size());
+
+  const float *src = cur_encoder_out->GetTensorData<float>();
+  float *dst = ans.GetTensorMutableData<float>();
+  for (int32_t i = 0; i != n; ++i) {
+    std::copy(src, src + cur_encoder_out_shape[1], dst);
+    dst += cur_encoder_out_shape[1];
+  }
+
+  return ans;
+}
+
 OnlineTransducerDecoderResult
 OnlineTransducerGreedySearchDecoder::GetEmptyResult() const {
   int32_t context_size = model_->ContextSize();
@@ -66,33 +90,33 @@ void OnlineTransducerGreedySearchDecoder::Decode(
     exit(-1);
   }
-  if (result->size() != 1) {
-    fprintf(stderr, "only batch size == 1 is implemented. Given: %d",
-            static_cast<int32_t>(result->size()));
-    exit(-1);
-  }
-
-  auto &hyp = (*result)[0].tokens;
-
-  int32_t num_frames = encoder_out_shape[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();
-  Ort::Value decoder_input = model_->BuildDecoderInput(hyp);
+  Ort::Value decoder_input = model_->BuildDecoderInput(*result);
   Ort::Value decoder_out = model_->RunDecoder(std::move(decoder_input));
   for (int32_t t = 0; t != num_frames; ++t) {
     Ort::Value cur_encoder_out = GetFrame(&encoder_out, t);
+    cur_encoder_out = Repeat(model_->Allocator(), &cur_encoder_out, batch_size);
     Ort::Value logit =
         model_->RunJoiner(std::move(cur_encoder_out), Clone(&decoder_out));
     const float *p_logit = logit.GetTensorData<float>();
-    auto y = static_cast<int32_t>(std::distance(
-        static_cast<const float *>(p_logit),
-        std::max_element(static_cast<const float *>(p_logit),
-                         static_cast<const float *>(p_logit) + vocab_size)));
-    if (y != 0) {
-      hyp.push_back(y);
-      decoder_input = model_->BuildDecoderInput(hyp);
+    bool emitted = false;
+    for (int32_t i = 0; i < batch_size; ++i, p_logit += vocab_size) {
+      auto y = static_cast<int32_t>(std::distance(
+          static_cast<const float *>(p_logit),
+          std::max_element(static_cast<const float *>(p_logit),
+                           static_cast<const float *>(p_logit) + vocab_size)));
+      if (y != 0) {
+        emitted = true;
+        (*result)[i].tokens.push_back(y);
+      }
+    }
+    if (emitted) {
+      decoder_input = model_->BuildDecoderInput(*result);
       decoder_out = model_->RunDecoder(std::move(decoder_input));
     }
   }
--- a/sherpa-onnx/csrc/online-transducer-model.h
查看文件 @8acc059
+++ b/sherpa-onnx/csrc/online-transducer-model.h
查看文件 @8acc059
@@ -13,6 +13,8 @@
 namespace sherpa_onnx {
+class OnlineTransducerDecoderResult;
+
 class OnlineTransducerModel {
  public:
   virtual ~OnlineTransducerModel() = default;
@@ -27,8 +29,8 @@ class OnlineTransducerModel {
    * @param states states[i] contains the state for the i-th utterance.
    * @return Return a single value representing the batched state.
    */
-  virtual Ort::Value StackStates(
-      const std::vector<Ort::Value> &states) const = 0;
+  virtual std::vector<Ort::Value> StackStates(
+      const std::vector<std::vector<Ort::Value>> &states) const = 0;
   /** Unstack a batch state into a list of individual states.
    *
@@ -37,7 +39,8 @@ class OnlineTransducerModel {
    * @param states A batched state.
    * @return ans[i] contains the state for the i-th utterance.
    */
-  virtual std::vector<Ort::Value> UnStackStates(Ort::Value states) const = 0;
+  virtual std::vector<std::vector<Ort::Value>> UnStackStates(
+      const std::vector<Ort::Value> &states) const = 0;
   /** Get the initial encoder states.
    *
@@ -58,7 +61,8 @@ class OnlineTransducerModel {
       Ort::Value features,
       std::vector<Ort::Value> &states) = 0;  // NOLINT
-  virtual Ort::Value BuildDecoderInput(const std::vector<int64_t> &hyp) = 0;
+  virtual Ort::Value BuildDecoderInput(
+      const std::vector<OnlineTransducerDecoderResult> &results) = 0;
   /** Run the decoder network.
    *
@@ -111,6 +115,7 @@ class OnlineTransducerModel {
   virtual int32_t VocabSize() const = 0;
   virtual int32_t SubsamplingFactor() const { return 4; }
+  virtual OrtAllocator *Allocator() = 0;
 };
 }  // namespace sherpa_onnx