Dynamic Parameters Workaround about cppadcodegen HOT 9 CLOSED

Given the sparsity patterns for matrix and the desired elements
https://coin-or.github.io/CppAD/doc/sparse_rc.htm

I would use row_major (or col_major) order to, in one pass of over the elements, determine which elements of the desired pattern are in the matrix pattern.

Perhaps it would be useful to write a general purpose function like this:
intersect_rc = intersection(first_rc, second_rc)
where _rc denotes a row and column sparsity pattern.

from cppadcodegen.

I have improved the error message for this case; see CppAD commit
coin-or/CppAD@30e4e4a

from cppadcodegen.

CppAD should give an error from a know soruce for this case and report the cause.
The problem is that the subset of elements requested must all be in the sparsity pattern.

This requirement is documented under the heading 'subset' on
https://coin-or.github.io/CppAD/doc/sparse_jac.htm

'All of the row, column pairs in subset must also appear in pattern ; i.e., they must be possibly non-zero. '

from cppadcodegen.

Okay so I correctly understood what the problem is. What's the best way to assure they are in the pattern?

from cppadcodegen.

Elements that are not in the pattern have value zero, so you do not need to calculate them.
The reason I did not handle this case is that it seems like a good cross check (that the pattern and subset of the pattern are consistent).

from cppadcodegen.

I guess what I'm asking is how to check if those values are zero and don't have to be added to my custom sparsity with setCustomSparseJacobianElements. So something like this:

// Jacobian sparsity
std::vector<size_t> rows;
std::vector<size_t> cols;
for (size_t row = 0; row < jacobian_t::RowsAtCompileTime; row++)
{
    for (size_t col = 0; col < jacobian_t::ColsAtCompileTime; col++)
    {
        if (/* (row, col) not already in pattern */)
        {
            rows.push_back(row);
            cols.push_back(col);
        }
    }
}

Or if there's a way to edit the current pattern so that I can remove the values that are not needed.

from cppadcodegen.

Okay with your help I ended up with something that works:

// sparsity pattern for the identity matrix
size_t nr = x.size();
size_t nc = x.size();
size_t nnz_in = x.size();
CppAD::sparse_rc<std::vector<size_t>> pattern_in(nr, nc, nnz_in);
for (size_t k = 0; k < nnz_in; k++)
{
    size_t r = k;
    size_t c = k;
    pattern_in.set(k, r, c);
}

CppAD::sparse_rc<std::vector<size_t>> pattern_out;
f_.for_jac_sparsity(pattern_in, false, false, false, pattern_out);

const std::vector<size_t> &row(pattern_out.row());
const std::vector<size_t> &col(pattern_out.col());

// reduced Jacobian sparsity
std::vector<size_t> rows;
std::vector<size_t> cols;

for (size_t i = 0; i < row.size(); i++)
{
    if (col[i] < desired_column_count)
    {
        rows.push_back(row[i]);
        cols.push_back(col[i]);
    }
}
cgen.setCustomSparseJacobianElements(rows, cols);

This is not actually faster than just computing the full dense Jacobian with all the parameter derivatives in it and then discarding the unused elements. Was this to be expected?

from cppadcodegen.

I am not sure how this all plays out in the context of CppADCodeGen, but in the context of CppAD,
it would be better to use the dynamic parameter option for the parameters that you are not taking
derivatives w.r.t.; see
https://coin-or.github.io/CppAD/doc/new_dynamic.htm

In addition, I only expect sparse versions to be faster when the matrices are large
and sparse. If you have such a case, in the context of CppAD, where the dense Jacobian is as fast as the sparse one, you could open an issue about it for CppAD; see
https://github.com/coin-or/CppAD/issues

from cppadcodegen.

I've been using CppAD parameters as well. The reason why I switched to code generation is the 10x speedup (even with the full Jacobian) that is absolutely necessary for my application.

The matrix is indeed pretty small with 20 independent parameters of which 10 are parameters, and 10 outputs. So the slower sparse Jacobian is not surprising.

I'll be very happy the day CppADCodegen impements parameters, but I understand it's a pretty big change and will take time.

Thanks again for your help.

from cppadcodegen.