lindahua / light-matrix Goto Github PK

implement and add this function to math_base.h

It is useful for various machine learning models, notably Latent Dirichlet Allocation.

Utilizing the begin and end iterators, a set of algorithms can be introduced

• count
• is_sorted
• sort
• partial_sort
• find_max
• find_min
• median

• const_marray and marray classes
• interface with cref_matrix and ref_matrix class

• gemv
• ger

• The ones has already been there
  • hypot (_hypot)
• rounding functions
  • round
  • trunc
• exponential and logarithm functions
  • exp2
  • log2
  • expm1
  • log1p
• hyperbolic functions (has equivalent math forms)
  • asinh
  • acosh
  • atanh
• error functions
  • erf
  • erfc

Reference: http://www.johndcook.com/math_h.html

as_col(A)    // map the entire matrix as a column view
as_row(A)   // map the entire matrix as a row view

appliable to all IDenseMatrix classes

Supports the following syntax

find_to(X > 0, it);  // it is an output iterator
find_to(X > 0, v);  // v is an instance of std::vector<index_t>
find_to(X > 0, I, J);  // I and J are both instances of index_seq

idx = find( x > 0 );  // returns an instance of class dense_col<index_t>

Y = select( A, idx );  // idx is any random access container
Y = select( A,  find(x > 0), colon(2, 3) );

• The find function should accept matrices with bool or mask type.
• The return type should be index_seq
• find function should be a simple wrapper of find_to.

• total reduction
• partial reduction (row/column)

• repeated rows/columns

Introduce a set of random matrix expressions

template<class Engine, class Distribution>  class random_expr;

Distr distr;
rand_mat( m, n,  distr, eng ); // ==> random_expr<Eng, Distr>
rand_mat( shape, distr, eng );

randu<T> ( m, n, eng );  // ==> random_expr<Eng, uniform_real_distribution<T> > 
randu<T> ( shape, eng ); 

randn<T> ( m, n, deg ); // ==> random_expr<Eng, normal_distribution<T> >
randn<T> ( shape, eng );

Notes:

• The value type should be inferrable from Distribution at compile-time.
• random_expr should implement IEWiseMatrix

• L2 blocking

dep/light-matrix/light_mat/matrix/matrix_layout.h: In member function ‘bool lmat::grid_layout<M, N>::is_percol_contiguous() const’:
dep/light-matrix/light_mat/matrix/matrix_layout.h:393:23: error: expression cannot be used as a function
    return m_rowstride() == 1;

The above line causes a compilation error, due to the trailing brackets. Removing them gets rid of the error.

• basic definitions
• meta-programming devices
• memory operations
• blocks --> arrays
• range

• Discrete
  • uniform_int_distr
  • bernoulli_distr
  • binomial_distr (a naive version)
  • geometric_distr (a naive version)
  • negative_binomial_distr (a naive version)
  • poisson_distr (a naive version)
  • discrete_distr (a naive version)
• Continuous
  • uniform_real_distr
  • normal_distr (use ICDF)
  • exponential_distr
  • gamma_distr
  • lognormal_distr

• arithmetics
• element-wise functions
• numerical comparison
• numerical logical operations on mask

• colview, colrange
• rowview, rowrange
• block subview

Introduce begin and end to IRegularMatrix

a.begin()  
a.end()
a.col_begin(j)
a.col_end(j)

begin(a);   // equivalent to a.begin()
end(a);      // equivalent to a.end()

Have both const and non-const versions

Argument wrapper allows some intermediate expression to be embedded into new expression.

This enables expression creation/transformation functions.

Syntax:

R = cond(Z > 0, X, Y);
R = cond(X > a, a, X);   // in MATLAB:  X(X > a) = a;

• The first argument must be a mask matrix
• The second and third argument can be either a matrix or a scalar (of the same type). At least one of them is a matrix.

The overall goal is to clarify the modules

• core ==> common
• improve accessor concept
• separate matrix classes, expression classes, and evaluation schemes
• generalize reduction interfaces

• transpose of vectors
• transpose of matrices

(no need to optimize at this point)

The syntax should be like

B = cast(A, type_<T>());

Here,

• A and B should have the same size.
• A and B can have different value types
• When A and B have the same value type, it simply calls evaluate_to

Convenient functions

to_f32(A);
to_f64(A);
to_bool(A); 
...

matrix comparison functions, which can be implemented using the all function.

is_equal(A, B);  // equivalent to all(A == B);
is_approx(A, B, tol); // equivalent to  all( abs(A - B) < tol );

Devices to inspect the structure of an expression

The implementation should be non-instrusive, relying on external template class with specializations.

// some declarations

template<class Expr> 
class  matexpr_typeinfo;

template<typename T, class Expr>
matexpr_typeinfo<Expr>  get_matexpr_type(const IMatrixXpr<Expr, T>& )
{
    return matexpr_typeinfo<Expr>();
}

// some usage: let a be an matrix expression instance

// return a type name (e.g. "unary_expr", "dense", etc) as std::string
get_matexpr_type( a ).name();  

// return a short string (e.g. "add(dense, sqr(dense))") as std::string
get_matexpr_type( a ).short_repr();  

• arithmetic calculation
• elementary function calculation
• comparison
• logical calculation

• common facilities
• matrix concepts and classes
• simd and math
• matrix expressions
• kernel-based evaluation
• linear algebra
• random number generation

Example:

g = _x * _x + exp(_x - _y)
g( 1 ) --> yields a scalar
g( arr ) --> yields an array

• architecture of the overall framework
• how to create new expressions
• kernel-based evaluation
• create new random distributions

full and partial reduction with the following functions

all(A, true);
all(A, false);
all(A);   // equivalent to all(A, true);

any(A, true);   
any(A, false);
any(A);   // equivalent to any(A, true); 

colwise_any(A, r, tf);
colwise_all(A, r, tf);

Syntax should be like:

f = poly(1, 2, 3) // f( x ) = x^2 + 2 * x + 3
f ( 1 ) --> produce a scalar
f ( arr ) --> produce an array through element-wise operation

Introduce a new module: random

Introduce a random stream concept

class IRandStream;

// a model based on std::random

template<class Engine> class std_randstream;

// a model based on SFMT

template<typename Kind> class sfmt_randstream;

// suppose s is an instance of IRandStream

s.rand_i32();  
s.rand_u32();
s.rand_pack( simd_kind );

s.rand_seq( nbytes, buf );

Generalized the IRange concept:

rgn.num();   // --> the number of indices (const)
rgn[i];   // --> the i-th index (const)

range_traits<R>::compile_time_num;

An IMatrixView Class:

class reindexed_matrix<T, RI, RJ>;

ct_rows == range_traits<RI>::compile_time_num;
ct_cols == range_traits<RJ>::compile_time_num;

C(rg0, rg1) returns an instance of reindexed_matrix, except when rg0 and rg1 are some special range.

• matrix base
  • concepts
  • meta-programming kits
  • properties
  • basic operations
• matrix classes
  • dense_matrix, dense_row, dense_col
  • ref_matrix, ref_row, ref_col
  • ref_matrix_ex

I think it would be good to allow building with -fno-exceptions, for better or worse this often is a requirement in practice. Would it be possible to make exception support optional?

• asum
• (scale)
• axpy
• dot
• nrm2
• rot

• gemv

• fill
• row-wise / column-wise fill

• Introduce a template class, mask_t
• Element-wise comparison should return matrices of mask_t
• mask_t is compatible with bool
• Element-wise logical operations on matrix of mask_t