Shark-ML is a C++ ML library with rich functionality. It also provides an API for linear algebra routines.  

There are four container classes for representing matrices and vectors in the Shark-ML library. Notice that the linear algebra functionality is declared in the remora namespace instead of the shark namespace, which is used for other routines.

The following code sample shows container classes that exist in the Shark-ML library, wherein the vector type is a dynamically sized array:

remora::vector<double> b(100, 1.0); // vector of size 100 and filled with 1.0

The compressed_vector type is a sparse array storing values in a compressed format.

The matrix type is a dynamically sized dense matrix, as shown in the following code snippet:

remora::matrix<double> C(2, 2); // 2x2 matrix

The compressed_matrix type is a sparse matrix storing values in a compressed format.

There are two main types of container initialization in the Shark-ML library.

We can initialize a container object with the constructor that takes the initializer list. The following code sample shows this:

remora::matrix<float> m_ones{{1, 1}, {1, 1}}; // 2x2 matrix

The second option is to wrap the existing C++ array into the container object and reuse its memory and values. The following code sample shows how to use the same array for the initialization of matrix and vector objects:

float data[]= {1,2,3,4};
remora::matrix<float> m(data, 2, 2);
remora::vector<float> v(data, 4);

Also, we can initialize values with direct access to the container elements, with the () operator. The following code sample shows how to set a value for matrix and vector objects:

remora::matrix<float> m(data, 2, 2);
m(0,0) = 3.14f;
remora::vector<float> v(data, 4);
v(0) = 3.14f;

The Shark-ML library implements linear algebra arithmetic operations through overloads of standard C++ arithmetic operators such as +, - and *. Some other operations such as the dot product are implemented as standalone functions.

The following code sample shows how to use arithmetic operations in the Shark-ML library:

remora::matrix<float> a(data, 2, 2);
remora::matrix<float> b(data, 2, 2);
auto c = a + b;
a -= b;
c = remora::prod(a,b);
c = a%b; // also dot product operation
c = a + 5;

We can use the following functions for partial access to the Shark ML containers:

• subrange (x,i,j): This function returns a sub-vector of x with the elements xi,…, xj−1.
• subrange (A,i,j,k,l): This function returns a sub-matrix of A with elements indicated by i,…, j−1 and k, …, l−1.
• row (A,k): This function returns the k^th row of A as a vector proxy.
• column (A,k): This function returns the k^th column of A as a vector proxy.
• rows (A,k,l): This function returns the rows k,…,l−1 of A as a matrix proxy.
• columns (A,k,l): This function returns the columns k,…, l−1 of A as a matrix proxy.

There is no broadcasting implementation in the Shark-ML library. Limited support of broadcasting exists only in the form of reduction functions (the set of functions that calculate one numeric value for a whole matrix or vector). There are two functions—the as_rows() and as_columns() function—that allow reduction operations to be performed independently on matrix rows or columns respectively. We can pass the result of these functions to any of the reduction functions. The following code sample shows how to perform summation reduction:

remora::matrix<float> m{{1, 2, 3, 4}, {5, 6, 7, 8}};
auto cols = remora::as_columns(m);
remora::sum(cols)

A different way to work with columns and rows independently is the use of partial access functions. The following code sample shows how to add the same vector to each of the matrix columns:

remora::vector<float> v{10, 10};
// Update matrix rows
for (size_t i = 0; i < m.size2(); ++i) {         
   remora::column(m, i) += v;
}