Access curr_grad_f, curr_grad_lag_x, curr_grad_lag_s, curr_exact_hessian in Ipopt::TNLP::intermediate_callback

[g-braeunlich]

Hello
From the docs
I understand, that I cannot directly use the return values given by the IpoptCalculatedQuantities::curr_* functions.
I first have to use some Resort* functions of an adapter.
What would be the Resort* functions corresponding to curr_grad_* and curr_exact_hessian?

[svigerske]

For gradients, ResortX should apply, but for entries that correspond to fixed variables, you will get the value of the variable instead of 0.
For the Hessian, I don't think that there is any function.

But what happens in ResortX isn't that complicated: So P_x_full_x_ is a permutation matrix that stores the mapping between indices in your TNLP and the NLP that you Ipopt works with internally. Unfortunately, it is private at the moment - fortunately you can just change code. There also seems to be a h_idx_map_ that could be useful.

[svigerske]

Ipopt 3.14 has some functions added to TNLP that allow to request the current iterate and various violations during intermediate_callback (or finalize_solution).
This gives access to the absolute values in curr_grad_lag_x.
curr_grad_lag_s should correspond to dual values of the constraints, so they should also be available.

I don't think I want to do anything to return the objective gradient or Hessian back to the user.