Suryanarayanan Manoj Sanu

Hi @mblondel , Do you know why this could happen?

@mblondel Is this going to be implemented ?. The reason why I need this because I have a blackbox solver from PETSc that I would like to use to solve...

Hey Martin, I will try to condense it into a MRE and post it here

@james-martens Hi James, Is it possible to create a custom loss for KFAC?

> Hey, > > Thank you for the implementation. > > From the guide, I saw that I have to register loss functions to be able to use K-FAC. For...

- Found other source: a. https://github.com/tianjuxue/jax-am/blob/1c321e58105e671d81444c0a8b97f37ea877567f/jax_am/fem/solver.py#L349. Add to end: ``` if use_petsc: A_fn.destroy() else: del A_fn del dofs del res_vec gc.collect() return sol ``` b. https://github.com/tianjuxue/jax-am/blob/1c321e58105e671d81444c0a8b97f37ea877567f/jax_am/fem/solver.py#L17 ``` # Delete PETSc...

I think we should make garbage collection as a standard practice when collaborating

Hey guys, Has there been any update on this aspect? It would be nice to have the optimizer-state exposed so that we can continue optimization!

``` def non_trace_force_stepping_fn(design, force_magnitues): u_init=0.0 for i in range(n_steps): force_magnitude = force_magnitudes[i] # Update value function problem_bcs = available_problem_bcs.ProblemBCsDatabase.get_problem('cantilever_2D', box_domain, force_magnitude=force_magnitude) fe_problem.neumann_bc_info = problem_bcs.neumann_bc_info # Solve the forward problem fwd_pred...

> Hi Matilda, > > thanks for the issue! > > For the points you are raising: do I understand correctly that you'd like `optim.update` to accept more keyword arguments...