cppadcg_det_minor.cpp

View page source

cppadcg Speed: Gradient of Determinant by Minor Expansion

Specifications

See link_det_minor .

PASS_JACOBIAN_TO_CODE_GEN

If this is one, the Jacobian of the determinant is the function passed to CppADCodeGen. In this case, the code_gen_fun function is used to calculate the Jacobian of the determinant. Otherwise, this flag is zero and the determinant function is passed to CppADCodeGen. In this case, the code_gen_fun jacobian is used to calculate the Jacobian of the determinant.

# define PASS_JACOBIAN_TO_CODE_GEN 1

Implementation

# include <cppad/speed/det_by_minor.hpp>
# include <cppad/speed/uniform_01.hpp>
# include <cppad/utility/vector.hpp>
# include <cppad/example/code_gen_fun.hpp>

# include <map>
extern std::map<std::string, bool> global_option;

namespace {
    //
    // typedefs
    typedef CppAD::cg::CG<double>     c_double;
    typedef CppAD::AD<c_double>      ac_double;
    typedef CppAD::vector<double>     d_vector;
    typedef CppAD::vector<ac_double> ac_vector;
    //
    // setup
    void setup(
        // inputs
        size_t size     ,
        // outputs
        code_gen_fun& fun )
    {  // optimization options
        std::string optimize_options =
            "no_conditional_skip no_compare_op no_print_for_op";
        //
        // object for computing determinant
        CppAD::det_by_minor<ac_double>   ac_det(size);
        //
        // number of independent variables
        size_t nx = size * size;
        //
        // choose a matrix
        CppAD::vector<double> matrix(nx);
        CppAD::uniform_01(nx, matrix);
        //
        // copy to independent variables
        ac_vector   ac_A(nx);
        for(size_t j = 0; j < nx; ++j)
            ac_A[j] = matrix[j];
        //
        // declare independent variables for function computation
        bool record_compare   = false;
        size_t abort_op_index = 0;
        CppAD::Independent(ac_A, abort_op_index, record_compare);
        //
        // AD computation of the determinant
        ac_vector ac_detA(1);
        ac_detA[0] = ac_det(ac_A);
        //
        // create function objects for f : A -> detA
        CppAD::ADFun<c_double>            c_f;
        c_f.Dependent(ac_A, ac_detA);
        if( global_option["optimize"] )
            c_f.optimize(optimize_options);
# if ! PASS_JACOBIAN_TO_CODE_GEN
        // f(x) is the determinant function
        code_gen_fun::evaluation_enum eval_jac = code_gen_fun::dense_enum;
        code_gen_fun f_tmp("det_minor", c_f, eval_jac);
        fun.swap(f_tmp);
# else
        CppAD::ADFun<ac_double, c_double> ac_f;
        ac_f = c_f.base2ad();
        //
        // declare independent variables for gradient computation
        CppAD::Independent(ac_A, abort_op_index, record_compare);
        //
        // vectors of reverse mode weights
        CppAD::vector<ac_double> ac_w(1);
        ac_w[0] = ac_double(1.0);
        //
        // AD computation of the gradient
        ac_vector ac_gradient(nx);
        ac_f.Forward(0, ac_A);
        ac_gradient = ac_f.Reverse(1, ac_w);
        //
        // create function objects for g : A -> det'( detA  )
        CppAD::ADFun<c_double> c_g;
        c_g.Dependent(ac_A, ac_gradient);
        if( global_option["optimize"] )
            c_g.optimize(optimize_options);
        // g(x) is the Jacobian of the determinant
        code_gen_fun g_tmp("det_minor", c_g);
        fun.swap(g_tmp);
# endif
    }
}

bool link_det_minor(
    const std::string&         job      ,
    size_t                     size     ,
    size_t                     repeat   ,
    CppAD::vector<double>     &matrix   ,
    CppAD::vector<double>     &gradient )
{  CPPAD_ASSERT_UNKNOWN( matrix.size() == size * size );
    CPPAD_ASSERT_UNKNOWN( gradient.size() == size * size );
    // --------------------------------------------------------------------
    // check global options
    const char* valid[] = { "onetape", "optimize"};
    size_t n_valid = sizeof(valid) / sizeof(valid[0]);
    typedef std::map<std::string, bool>::iterator iterator;
    //
    for(iterator itr=global_option.begin(); itr!=global_option.end(); ++itr)
    {  if( itr->second )
        {  bool ok = false;
            for(size_t i = 0; i < n_valid; i++)
                ok |= itr->first == valid[i];
            if( ! ok )
                return false;
        }
    }
    // --------------------------------------------------------------------
    //
    // function object mapping matrix to gradient of determinant
    static code_gen_fun static_fun;
    //
    // size corresponding static_fun
    static size_t static_size = 0;
    //
    // number of independent variables
    size_t nx = size * size;
    //
    // onetape
    bool onetape = global_option["onetape"];
    // ----------------------------------------------------------------------
    if( job == "setup" )
    {  if( onetape )
        {  setup(size, static_fun);
            static_size = size;
        }
        else
        {  static_size = 0;
        }
        return true;
    }
    if( job ==  "teardown" )
    {  code_gen_fun fun;
        static_fun.swap(fun);
        return true;
    }
    // -----------------------------------------------------------------------
    CPPAD_ASSERT_UNKNOWN( job == "run" );
    if( onetape ) while(repeat--)
    {  // use if before assert to avoid warning that static_size is not used
        if( size != static_size )
        {  CPPAD_ASSERT_UNKNOWN( size == static_size );
        }

        // get next matrix
        CppAD::uniform_01(nx, matrix);

        // evaluate the gradient
# if PASS_JACOBIAN_TO_CODE_GEN
        gradient = static_fun(matrix);
# else
        gradient = static_fun.jacobian(matrix);
# endif
    }
    else while(repeat--)
    {  setup(size, static_fun);
        static_size = size;

        // get next matrix
        CppAD::uniform_01(nx, matrix);

        // evaluate the gradient
# if PASS_JACOBIAN_TO_CODE_GEN
        gradient = static_fun(matrix);
# else
        gradient = static_fun.jacobian(matrix);
# endif
    }
    return true;
}