From 96a32415ab43377cf1575bd3f4f2980f58028209 Mon Sep 17 00:00:00 2001 From: Determinant Date: Fri, 14 Aug 2015 11:51:42 +0800 Subject: add implementation for kaldi io (by ymz) --- kaldi_io/src/kaldi/matrix/kaldi-vector.h | 585 +++++++++++++++++++++++++++++++ 1 file changed, 585 insertions(+) create mode 100644 kaldi_io/src/kaldi/matrix/kaldi-vector.h (limited to 'kaldi_io/src/kaldi/matrix/kaldi-vector.h') diff --git a/kaldi_io/src/kaldi/matrix/kaldi-vector.h b/kaldi_io/src/kaldi/matrix/kaldi-vector.h new file mode 100644 index 0000000..2b3395b --- /dev/null +++ b/kaldi_io/src/kaldi/matrix/kaldi-vector.h @@ -0,0 +1,585 @@ +// matrix/kaldi-vector.h + +// Copyright 2009-2012 Ondrej Glembek; Microsoft Corporation; Lukas Burget; +// Saarland University (Author: Arnab Ghoshal); +// Ariya Rastrow; Petr Schwarz; Yanmin Qian; +// Karel Vesely; Go Vivace Inc.; Arnab Ghoshal +// Wei Shi; + +// See ../../COPYING for clarification regarding multiple authors +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED +// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE, +// MERCHANTABLITY OR NON-INFRINGEMENT. +// See the Apache 2 License for the specific language governing permissions and +// limitations under the License. + +#ifndef KALDI_MATRIX_KALDI_VECTOR_H_ +#define KALDI_MATRIX_KALDI_VECTOR_H_ 1 + +#include "matrix/matrix-common.h" + +namespace kaldi { + +/// \addtogroup matrix_group +/// @{ + +/// Provides a vector abstraction class. +/// This class provides a way to work with vectors in kaldi. +/// It encapsulates basic operations and memory optimizations. +template +class VectorBase { + public: + /// Set vector to all zeros. + void SetZero(); + + /// Returns true if matrix is all zeros. + bool IsZero(Real cutoff = 1.0e-06) const; // replace magic number + + /// Set all members of a vector to a specified value. + void Set(Real f); + + /// Set vector to random normally-distributed noise. + void SetRandn(); + + /// This function returns a random index into this vector, + /// chosen with probability proportional to the corresponding + /// element. Requires that this->Min() >= 0 and this->Sum() > 0. + MatrixIndexT RandCategorical() const; + + /// Returns the dimension of the vector. + inline MatrixIndexT Dim() const { return dim_; } + + /// Returns the size in memory of the vector, in bytes. + inline MatrixIndexT SizeInBytes() const { return (dim_*sizeof(Real)); } + + /// Returns a pointer to the start of the vector's data. + inline Real* Data() { return data_; } + + /// Returns a pointer to the start of the vector's data (const). + inline const Real* Data() const { return data_; } + + /// Indexing operator (const). + inline Real operator() (MatrixIndexT i) const { + KALDI_PARANOID_ASSERT(static_cast(i) < + static_cast(dim_)); + return *(data_ + i); + } + + /// Indexing operator (non-const). + inline Real & operator() (MatrixIndexT i) { + KALDI_PARANOID_ASSERT(static_cast(i) < + static_cast(dim_)); + return *(data_ + i); + } + + /** @brief Returns a sub-vector of a vector (a range of elements). + * @param o [in] Origin, 0 < o < Dim() + * @param l [in] Length 0 < l < Dim()-o + * @return A SubVector object that aliases the data of the Vector object. + * See @c SubVector class for details */ + SubVector Range(const MatrixIndexT o, const MatrixIndexT l) { + return SubVector(*this, o, l); + } + + /** @brief Returns a const sub-vector of a vector (a range of elements). + * @param o [in] Origin, 0 < o < Dim() + * @param l [in] Length 0 < l < Dim()-o + * @return A SubVector object that aliases the data of the Vector object. + * See @c SubVector class for details */ + const SubVector Range(const MatrixIndexT o, + const MatrixIndexT l) const { + return SubVector(*this, o, l); + } + + /// Copy data from another vector (must match own size). + void CopyFromVec(const VectorBase &v); + + /// Copy data from a SpMatrix or TpMatrix (must match own size). + template + void CopyFromPacked(const PackedMatrix &M); + + /// Copy data from another vector of different type (double vs. float) + template + void CopyFromVec(const VectorBase &v); + + /// Copy from CuVector. This is defined in ../cudamatrix/cu-vector.h + template + void CopyFromVec(const CuVectorBase &v); + + + /// Apply natural log to all elements. Throw if any element of + /// the vector is negative (but doesn't complain about zero; the + /// log will be -infinity + void ApplyLog(); + + /// Apply natural log to another vector and put result in *this. + void ApplyLogAndCopy(const VectorBase &v); + + /// Apply exponential to each value in vector. + void ApplyExp(); + + /// Take absolute value of each of the elements + void ApplyAbs(); + + /// Applies floor to all elements. Returns number of elements floored. + MatrixIndexT ApplyFloor(Real floor_val); + + /// Applies ceiling to all elements. Returns number of elements changed. + MatrixIndexT ApplyCeiling(Real ceil_val); + + /// Applies floor to all elements. Returns number of elements floored. + MatrixIndexT ApplyFloor(const VectorBase &floor_vec); + + /// Apply soft-max to vector and return normalizer (log sum of exponentials). + /// This is the same as: \f$ x(i) = exp(x(i)) / \sum_i exp(x(i)) \f$ + Real ApplySoftMax(); + + /// Sets each element of *this to the tanh of the corresponding element of "src". + void Tanh(const VectorBase &src); + + /// Sets each element of *this to the sigmoid function of the corresponding + /// element of "src". + void Sigmoid(const VectorBase &src); + + /// Take all elements of vector to a power. + void ApplyPow(Real power); + + /// Take the absolute value of all elements of a vector to a power. + /// Include the sign of the input element if include_sign == true. + /// If power is negative and the input value is zero, the output is set zero. + void ApplyPowAbs(Real power, bool include_sign=false); + + /// Compute the p-th norm of the vector. + Real Norm(Real p) const; + + /// Returns true if ((*this)-other).Norm(2.0) <= tol * (*this).Norm(2.0). + bool ApproxEqual(const VectorBase &other, float tol = 0.01) const; + + /// Invert all elements. + void InvertElements(); + + /// Add vector : *this = *this + alpha * rv (with casting between floats and + /// doubles) + template + void AddVec(const Real alpha, const VectorBase &v); + + /// Add vector : *this = *this + alpha * rv^2 [element-wise squaring]. + void AddVec2(const Real alpha, const VectorBase &v); + + /// Add vector : *this = *this + alpha * rv^2 [element-wise squaring], + /// with casting between floats and doubles. + template + void AddVec2(const Real alpha, const VectorBase &v); + + /// Add matrix times vector : this <-- beta*this + alpha*M*v. + /// Calls BLAS GEMV. + void AddMatVec(const Real alpha, const MatrixBase &M, + const MatrixTransposeType trans, const VectorBase &v, + const Real beta); // **beta previously defaulted to 0.0** + + /// This is as AddMatVec, except optimized for where v contains a lot + /// of zeros. + void AddMatSvec(const Real alpha, const MatrixBase &M, + const MatrixTransposeType trans, const VectorBase &v, + const Real beta); // **beta previously defaulted to 0.0** + + + /// Add symmetric positive definite matrix times vector: + /// this <-- beta*this + alpha*M*v. Calls BLAS SPMV. + void AddSpVec(const Real alpha, const SpMatrix &M, + const VectorBase &v, const Real beta); // **beta previously defaulted to 0.0** + + /// Add triangular matrix times vector: this <-- beta*this + alpha*M*v. + /// Works even if rv == *this. + void AddTpVec(const Real alpha, const TpMatrix &M, + const MatrixTransposeType trans, const VectorBase &v, + const Real beta); // **beta previously defaulted to 0.0** + + /// Set each element to y = (x == orig ? changed : x). + void ReplaceValue(Real orig, Real changed); + + /// Multipy element-by-element by another vector. + void MulElements(const VectorBase &v); + /// Multipy element-by-element by another vector of different type. + template + void MulElements(const VectorBase &v); + + /// Divide element-by-element by a vector. + void DivElements(const VectorBase &v); + /// Divide element-by-element by a vector of different type. + template + void DivElements(const VectorBase &v); + + /// Add a constant to each element of a vector. + void Add(Real c); + + /// Add element-by-element product of vectlrs: + // this <-- alpha * v .* r + beta*this . + void AddVecVec(Real alpha, const VectorBase &v, + const VectorBase &r, Real beta); + + /// Add element-by-element quotient of two vectors. + /// this <---- alpha*v/r + beta*this + void AddVecDivVec(Real alpha, const VectorBase &v, + const VectorBase &r, Real beta); + + /// Multiplies all elements by this constant. + void Scale(Real alpha); + + /// Multiplies this vector by lower-triangular marix: *this <-- *this *M + void MulTp(const TpMatrix &M, const MatrixTransposeType trans); + + /// If trans == kNoTrans, solves M x = b, where b is the value of *this at input + /// and x is the value of *this at output. + /// If trans == kTrans, solves M' x = b. + /// Does not test for M being singular or near-singular, so test it before + /// calling this routine. + void Solve(const TpMatrix &M, const MatrixTransposeType trans); + + /// Performs a row stack of the matrix M + void CopyRowsFromMat(const MatrixBase &M); + template + void CopyRowsFromMat(const MatrixBase &M); + + /// The following is implemented in ../cudamatrix/cu-matrix.cc + void CopyRowsFromMat(const CuMatrixBase &M); + + /// Performs a column stack of the matrix M + void CopyColsFromMat(const MatrixBase &M); + + /// Extracts a row of the matrix M. Could also do this with + /// this->Copy(M[row]). + void CopyRowFromMat(const MatrixBase &M, MatrixIndexT row); + /// Extracts a row of the matrix M with type conversion. + template + void CopyRowFromMat(const MatrixBase &M, MatrixIndexT row); + + /// Extracts a row of the symmetric matrix S. + template + void CopyRowFromSp(const SpMatrix &S, MatrixIndexT row); + + /// Extracts a column of the matrix M. + template + void CopyColFromMat(const MatrixBase &M , MatrixIndexT col); + + /// Extracts the diagonal of the matrix M. + void CopyDiagFromMat(const MatrixBase &M); + + /// Extracts the diagonal of a packed matrix M; works for Sp or Tp. + void CopyDiagFromPacked(const PackedMatrix &M); + + + /// Extracts the diagonal of a symmetric matrix. + inline void CopyDiagFromSp(const SpMatrix &M) { CopyDiagFromPacked(M); } + + /// Extracts the diagonal of a triangular matrix. + inline void CopyDiagFromTp(const TpMatrix &M) { CopyDiagFromPacked(M); } + + /// Returns the maximum value of any element, or -infinity for the empty vector. + Real Max() const; + + /// Returns the maximum value of any element, and the associated index. + /// Error if vector is empty. + Real Max(MatrixIndexT *index) const; + + /// Returns the minimum value of any element, or +infinity for the empty vector. + Real Min() const; + + /// Returns the minimum value of any element, and the associated index. + /// Error if vector is empty. + Real Min(MatrixIndexT *index) const; + + /// Returns sum of the elements + Real Sum() const; + + /// Returns sum of the logs of the elements. More efficient than + /// just taking log of each. Will return NaN if any elements are + /// negative. + Real SumLog() const; + + /// Does *this = alpha * (sum of rows of M) + beta * *this. + void AddRowSumMat(Real alpha, const MatrixBase &M, Real beta = 1.0); + + /// Does *this = alpha * (sum of columns of M) + beta * *this. + void AddColSumMat(Real alpha, const MatrixBase &M, Real beta = 1.0); + + /// Add the diagonal of a matrix times itself: + /// *this = diag(M M^T) + beta * *this (if trans == kNoTrans), or + /// *this = diag(M^T M) + beta * *this (if trans == kTrans). + void AddDiagMat2(Real alpha, const MatrixBase &M, + MatrixTransposeType trans = kNoTrans, Real beta = 1.0); + + /// Add the diagonal of a matrix product: *this = diag(M N), assuming the + /// "trans" arguments are both kNoTrans; for transpose arguments, it behaves + /// as you would expect. + void AddDiagMatMat(Real alpha, const MatrixBase &M, MatrixTransposeType transM, + const MatrixBase &N, MatrixTransposeType transN, + Real beta = 1.0); + + /// Returns log(sum(exp())) without exp overflow + /// If prune > 0.0, ignores terms less than the max - prune. + /// [Note: in future, if prune = 0.0, it will take the max. + /// For now, use -1 if you don't want it to prune.] + Real LogSumExp(Real prune = -1.0) const; + + /// Reads from C++ stream (option to add to existing contents). + /// Throws exception on failure + void Read(std::istream & in, bool binary, bool add = false); + + /// Writes to C++ stream (option to write in binary). + void Write(std::ostream &Out, bool binary) const; + + friend class VectorBase; + friend class VectorBase; + friend class CuVectorBase; + friend class CuVector; + protected: + /// Destructor; does not deallocate memory, this is handled by child classes. + /// This destructor is protected so this object so this object can only be + /// deleted via a child. + ~VectorBase() {} + + /// Empty initializer, corresponds to vector of zero size. + explicit VectorBase(): data_(NULL), dim_(0) { + KALDI_ASSERT_IS_FLOATING_TYPE(Real); + } + +// Took this out since it is not currently used, and it is possible to create +// objects where the allocated memory is not the same size as dim_ : Arnab +// /// Initializer from a pointer and a size; keeps the pointer internally +// /// (ownership or non-ownership depends on the child class). +// explicit VectorBase(Real* data, MatrixIndexT dim) +// : data_(data), dim_(dim) {} + + // Arnab : made this protected since it is unsafe too. + /// Load data into the vector: sz must match own size. + void CopyFromPtr(const Real* Data, MatrixIndexT sz); + + /// data memory area + Real* data_; + /// dimension of vector + MatrixIndexT dim_; + KALDI_DISALLOW_COPY_AND_ASSIGN(VectorBase); +}; // class VectorBase + +/** @brief A class representing a vector. + * + * This class provides a way to work with vectors in kaldi. + * It encapsulates basic operations and memory optimizations. */ +template +class Vector: public VectorBase { + public: + /// Constructor that takes no arguments. Initializes to empty. + Vector(): VectorBase() {} + + /// Constructor with specific size. Sets to all-zero by default + /// if set_zero == false, memory contents are undefined. + explicit Vector(const MatrixIndexT s, + MatrixResizeType resize_type = kSetZero) + : VectorBase() { Resize(s, resize_type); } + + /// Copy constructor from CUDA vector + /// This is defined in ../cudamatrix/cu-vector.h + template + explicit Vector(const CuVectorBase &cu); + + /// Copy constructor. The need for this is controversial. + Vector(const Vector &v) : VectorBase() { // (cannot be explicit) + Resize(v.Dim(), kUndefined); + this->CopyFromVec(v); + } + + /// Copy-constructor from base-class, needed to copy from SubVector. + explicit Vector(const VectorBase &v) : VectorBase() { + Resize(v.Dim(), kUndefined); + this->CopyFromVec(v); + } + + /// Type conversion constructor. + template + explicit Vector(const VectorBase &v): VectorBase() { + Resize(v.Dim(), kUndefined); + this->CopyFromVec(v); + } + +// Took this out since it is unsafe : Arnab +// /// Constructor from a pointer and a size; copies the data to a location +// /// it owns. +// Vector(const Real* Data, const MatrixIndexT s): VectorBase() { +// Resize(s); + // CopyFromPtr(Data, s); +// } + + + /// Swaps the contents of *this and *other. Shallow swap. + void Swap(Vector *other); + + /// Destructor. Deallocates memory. + ~Vector() { Destroy(); } + + /// Read function using C++ streams. Can also add to existing contents + /// of matrix. + void Read(std::istream & in, bool binary, bool add = false); + + /// Set vector to a specified size (can be zero). + /// The value of the new data depends on resize_type: + /// -if kSetZero, the new data will be zero + /// -if kUndefined, the new data will be undefined + /// -if kCopyData, the new data will be the same as the old data in any + /// shared positions, and zero elsewhere. + /// This function takes time proportional to the number of data elements. + void Resize(MatrixIndexT length, MatrixResizeType resize_type = kSetZero); + + /// Remove one element and shifts later elements down. + void RemoveElement(MatrixIndexT i); + + /// Assignment operator, protected so it can only be used by std::vector + Vector &operator = (const Vector &other) { + Resize(other.Dim(), kUndefined); + this->CopyFromVec(other); + return *this; + } + + /// Assignment operator that takes VectorBase. + Vector &operator = (const VectorBase &other) { + Resize(other.Dim(), kUndefined); + this->CopyFromVec(other); + return *this; + } + private: + /// Init assumes the current contents of the class are invalid (i.e. junk or + /// has already been freed), and it sets the vector to newly allocated memory + /// with the specified dimension. dim == 0 is acceptable. The memory contents + /// pointed to by data_ will be undefined. + void Init(const MatrixIndexT dim); + + /// Destroy function, called internally. + void Destroy(); + +}; + + +/// Represents a non-allocating general vector which can be defined +/// as a sub-vector of higher-level vector [or as the row of a matrix]. +template +class SubVector : public VectorBase { + public: + /// Constructor from a Vector or SubVector. + /// SubVectors are not const-safe and it's very hard to make them + /// so for now we just give up. This function contains const_cast. + SubVector(const VectorBase &t, const MatrixIndexT origin, + const MatrixIndexT length) : VectorBase() { + // following assert equiv to origin>=0 && length>=0 && + // origin+length <= rt.dim_ + KALDI_ASSERT(static_cast(origin)+ + static_cast(length) <= + static_cast(t.Dim())); + VectorBase::data_ = const_cast (t.Data()+origin); + VectorBase::dim_ = length; + } + + /// This constructor initializes the vector to point at the contents + /// of this packed matrix (SpMatrix or TpMatrix). + SubVector(const PackedMatrix &M) { + VectorBase::data_ = const_cast (M.Data()); + VectorBase::dim_ = (M.NumRows()*(M.NumRows()+1))/2; + } + + /// Copy constructor + SubVector(const SubVector &other) : VectorBase () { + // this copy constructor needed for Range() to work in base class. + VectorBase::data_ = other.data_; + VectorBase::dim_ = other.dim_; + } + + /// Constructor from a pointer to memory and a length. Keeps a pointer + /// to the data but does not take ownership (will never delete). + SubVector(Real *data, MatrixIndexT length) : VectorBase () { + VectorBase::data_ = data; + VectorBase::dim_ = length; + } + + + /// This operation does not preserve const-ness, so be careful. + SubVector(const MatrixBase &matrix, MatrixIndexT row) { + VectorBase::data_ = const_cast(matrix.RowData(row)); + VectorBase::dim_ = matrix.NumCols(); + } + + ~SubVector() {} ///< Destructor (does nothing; no pointers are owned here). + + private: + /// Disallow assignment operator. + SubVector & operator = (const SubVector &other) {} +}; + +/// @} end of "addtogroup matrix_group" +/// \addtogroup matrix_funcs_io +/// @{ +/// Output to a C++ stream. Non-binary by default (use Write for +/// binary output). +template +std::ostream & operator << (std::ostream & out, const VectorBase & v); + +/// Input from a C++ stream. Will automatically read text or +/// binary data from the stream. +template +std::istream & operator >> (std::istream & in, VectorBase & v); + +/// Input from a C++ stream. Will automatically read text or +/// binary data from the stream. +template +std::istream & operator >> (std::istream & in, Vector & v); +/// @} end of \addtogroup matrix_funcs_io + +/// \addtogroup matrix_funcs_scalar +/// @{ + + +template +bool ApproxEqual(const VectorBase &a, + const VectorBase &b, Real tol = 0.01) { + return a.ApproxEqual(b, tol); +} + +template +inline void AssertEqual(VectorBase &a, VectorBase &b, + float tol = 0.01) { + KALDI_ASSERT(a.ApproxEqual(b, tol)); +} + + +/// Returns dot product between v1 and v2. +template +Real VecVec(const VectorBase &v1, const VectorBase &v2); + +template +Real VecVec(const VectorBase &v1, const VectorBase &v2); + + +/// Returns \f$ v_1^T M v_2 \f$ . +/// Not as efficient as it could be where v1 == v2. +template +Real VecMatVec(const VectorBase &v1, const MatrixBase &M, + const VectorBase &v2); + +/// @} End of "addtogroup matrix_funcs_scalar" + + +} // namespace kaldi + +// we need to include the implementation +#include "matrix/kaldi-vector-inl.h" + + + +#endif // KALDI_MATRIX_KALDI_VECTOR_H_ + -- cgit v1.2.3