Using evalJacobianAndHessian helper for variables larger than 3

[Manushanker]

Hi @MaxSagebaum

I am using the EH::evalJacobianAndHessian helper to compute the Jacobian and Hessian of a user defined function and the code works perfectly when the number of variables is 3. I would like to extend it for arbitrary number of variables. In the documentation it is mentioned that the default size of the "HessianComputationType" is 4 and so, I modified the hessianComputationType as "using codi::EvaluationHelper::HessianComputationType = RealForwardGen<RealForwardVec<5>, Direction<RealForwardVec<5>, 5> >" in evaluationHelper.hpp file but faced runtime errors. Can you please let me know how to use the evalJacodianHessian helper for variables greater than three.

Thanks,

Manushanker

[Manushanker]

Hi @MaxSagebaum

As mentioned in my previous comment the "evalJacobianAndHessian" helper when using the "HessianComputationType" computes Hessian and jacobian matrix up to four variables. Can you please let me know how to compute the derivatives when the number of variables is greater than 4?

Here is my code
auto Effieldjac = EH::createJacobian(1, arg.size()); // arg is a std vector with a list of variables
auto Effieldhes = EH::createHessian(1, arg.size());
EH::evalJacobianAndHessian(codiDotWithtestwrapper<EH::HessianComputationType>, arg, 1, Effieldjac, Effieldhes);
// codiDotWithtestwrapper is a wrapper function
template<l>
void codiDotWithtestwrapper(std::vector<> const x, std::vector<>& y)
{
func(&x[0], y[0]);
}
// func is a user defined function
template<>
void func(T const* arg, T& FFEfield)
{

}

Thanks,

Manushanker

[MaxSagebaum]

The 4 in the definition of the HessisonComputationType defines the vector mode of AD. This value is independent from the number of inputs and outputs. So arg.size() can be arbitrary without changing your code.

Can you be more specific on what your problem is?

[Manushanker]

Thanks for the clarifications! When I use the arg.size() >4 the library does not compute the Jacobian and Hessian matrix. Currently, the code works only for certain examples with 4 variables. When I ran the code for more complex example with 4 variables the Hessian and Jacobian matrix values are all zero. A few months back I sent you a code and you helped to resolve it. Now, I integrated the CoDiPack with more complicated in-house code. If needed I can send you my code for review.
Regards,

Manushanker

[MaxSagebaum]

I created an example, that computes a Jacobian and Hessian for more than 4 variables, using the EvaluationHelper::HessianComputationType. The derivatives are correct and there is no problem with the algorithm by itself.

If you want you can send me the code and I can have a quick look. It would be nice if you could describe what approaches you have already taken to find the problem.

On possible cause might be, that the function you provide to the evalJacobianAndHessian method, is not stateless or has side effects. CoDiPack calls this function multiple times and the behavior is undefined if the function does not compute the same values in each call.

evalHelperExample.cpp.txt

[MaxSagebaum]

There is a segmentation fault in the code you send me. In febi.h:956 the parameter ifile is a null pointer when the function is called a second time.

Please check your code for memory errors. You can do this with valgrind or compile with -fsanitize=address (gcc only).

In the discussion we had earlier this year, you already mentioned, that you would like to avoid reading the input files multiple times. I showed you how to use the TapeHelper to avoid this. You might want to switch to this solution.

[Manushanker]

Thanks for looking at my code. Yes, I want to avoid calling derivative function multiple times. As you suggested, I am using the TapeHelper to compute the Jacobian and Hessian. I am currently facing runtime errors and trying to resolve it. Will update you its status once I resolve it.