Lattice decoder (#10)

kaldi-decoder/csrc/CMakeLists.txt

@@ -5,9 +5,11 @@ set(srcs
   decodable-ctc.cc
   eigen.cc
   faster-decoder.cc
+  lattice-faster-decoder.cc
+  lattice-simple-decoder.cc
+  simple-decoder.cc
 )
 
-
 add_library(kaldi-decoder-core ${srcs})
 
 target_link_libraries(kaldi-decoder-core PUBLIC kaldifst_core)

kaldi-decoder/csrc/decodable-ctc.cc

@@ -8,26 +8,27 @@
 
 namespace kaldi_decoder {
 
-DecodableCtc::DecodableCtc(const FloatMatrix &log_probs)
-    : log_probs_(log_probs) {
+DecodableCtc::DecodableCtc(const FloatMatrix &log_probs, int32_t offset /*= 0*/)
+    : log_probs_(log_probs), offset_(offset) {
   p_ = &log_probs_(0, 0);
   num_rows_ = log_probs_.rows();
   num_cols_ = log_probs_.cols();
 }
 
-DecodableCtc::DecodableCtc(const float *p, int32_t num_rows, int32_t num_cols)
-    : p_(p), num_rows_(num_rows), num_cols_(num_cols) {}
+DecodableCtc::DecodableCtc(const float *p, int32_t num_rows, int32_t num_cols,
+                           int32_t offset /*= 0*/)
+    : p_(p), num_rows_(num_rows), num_cols_(num_cols), offset_(offset) {}
 
 float DecodableCtc::LogLikelihood(int32_t frame, int32_t index) {
   // Note: We need to use index - 1 here since
   // all the input labels of the H are incremented during graph
   // construction
   assert(index >= 1);
 
-  return *(p_ + frame * num_cols_ + index - 1);
+  return *(p_ + (frame - offset_) * num_cols_ + index - 1);
 }
 
-int32_t DecodableCtc::NumFramesReady() const { return num_rows_; }
+int32_t DecodableCtc::NumFramesReady() const { return offset_ + num_rows_; }
 
 int32_t DecodableCtc::NumIndices() const { return num_cols_; }
 

kaldi-decoder/csrc/decodable-ctc.h

@@ -13,14 +13,15 @@ namespace kaldi_decoder {
 class DecodableCtc : public DecodableInterface {
  public:
   // It copies the input log_probs
-  explicit DecodableCtc(const FloatMatrix &log_probs);
+  explicit DecodableCtc(const FloatMatrix &log_probs, int32_t offset = 0);
 
   // It shares the memory with the input array.
   //
   // @param p Pointer to a 2-d array of shape (num_rows, num_cols).
   //          The array should be kept alive as long as this object is still
   //          alive.
-  DecodableCtc(const float *p, int32_t num_rows, int32_t num_cols);
+  DecodableCtc(const float *p, int32_t num_rows, int32_t num_cols,
+               int32_t offset = 0);
 
   float LogLikelihood(int32_t frame, int32_t index) override;
 
@@ -38,6 +39,7 @@ class DecodableCtc : public DecodableInterface {
   const float *p_ = nullptr;  // pointer to a 2-d array
   int32_t num_rows_;          // number of rows in the 2-d array
   int32_t num_cols_;          // number of cols in the 2-d array
+  int32_t offset_ = 0;
 };
 
 }  // namespace kaldi_decoder

kaldi-decoder/csrc/kaldi-math.h

@@ -0,0 +1,48 @@
+// kaldi-decoder/csrc/kaldi-math.h
+
+// Copyright 2009-2011  Ondrej Glembek;  Microsoft Corporation;  Yanmin Qian;
+//                      Jan Silovsky;  Saarland University
+//                2023  Xiaomi Corporation
+
+// this file is copied and modified from
+// kaldi/src/base/kaldi-math.h
+#ifndef KALDI_DECODER_CSRC_KALDI_MATH_H_
+#define KALDI_DECODER_CSRC_KALDI_MATH_H_
+
+#include <cmath>
+#include <limits>
+
+#ifndef DBL_EPSILON
+#define DBL_EPSILON 2.2204460492503131e-16
+#endif
+
+#ifndef FLT_EPSILON
+#define FLT_EPSILON 1.19209290e-7f
+#endif
+
+// M_LOG_2PI =  log(2*pi)
+#ifndef M_LOG_2PI
+#define M_LOG_2PI 1.8378770664093454835606594728112
+#endif
+
+#define KALDI_ISINF std::isinf
+#define KALDI_ISNAN std::isnan
+
+#include "kaldi-decoder/csrc/log.h"
+
+namespace kaldi_decoder {
+
+/// return abs(a - b) <= relative_tolerance * (abs(a)+abs(b)).
+static inline bool ApproxEqual(float a, float b,
+                               float relative_tolerance = 0.001) {
+  // a==b handles infinities.
+  if (a == b) return true;
+  float diff = std::abs(a - b);
+  if (diff == std::numeric_limits<float>::infinity() || diff != diff)
+    return false;  // diff is +inf or nan.
+  return (diff <= relative_tolerance * (std::abs(a) + std::abs(b)));
+}
+
+}  // namespace kaldi_decoder
+
+#endif  // KALDI_DECODER_CSRC_KALDI_MATH_H_

kaldi-decoder/csrc/lattice-faster-decoder.cc

@@ -0,0 +1,13 @@
+// kaldi-decoder/csrc/lattice-faster-decoder.cc
+
+// Copyright 2009-2012  Microsoft Corporation  Mirko Hannemann
+//           2013-2018  Johns Hopkins University (Author: Daniel Povey)
+//                2014  Guoguo Chen
+//                2018  Zhehuai Chen
+// Copyright (c)  2023  Xiaomi Corporation
+
+// this file is copied and modified from
+// kaldi/src/decoder/lattice-faster-decoder.cc
+
+#include "kaldi-decoder/csrc/lattice-faster-decoder.h"
+namespace kaldi_decoder {}