why optimization doesn't work

[EXing]

code from #596

Start theseus iteration:
Nonlinear optimizer. Iteration: 0. Error: 8258.4384765625
Nonlinear optimizer. Iteration: 1. Error: 8258.4384765625
End theseus iteration.

why was the optimization finished directly?

[EXing]

how can I find out which end condition is triggered?

[mhmukadam]

You could setup breakpoints in this function to see which condition is triggered. The error stays the same between the two iterations so likely it's the first one (error abs change is less than threshold).

[luisenp]

@EXing The TheseusLayer.forward() method returns and info object that has status codes for termination conditions. When converging in one step a common cause is 0 gradients leading to immediate (incorrect) convergence.

[EXing]

# For now, step size is combined with delta. If we add more sophisticated
# line search, will probably need to pass it separately, or compute inside.
err, all_rejected = self._step(
    delta * steps_tensor,
    info.last_err,
    converged_indices,
    force_update,
    delta_forced_to_gn=_last_implicit_diff_step,
    **kwargs,
)  # err is shape (batch_size,)

all_rejected: True
converged_indices: [False]
force_update: False
_last_implicit_diff_step: False

Find out all update was rejected, any help?

[EXing]

reject_indices = self._complete_step(delta, err, previous_err, **kwargs)

    def _complete_step(
        self,
        delta: torch.Tensor,
        new_err: torch.Tensor,
        previous_err: torch.Tensor,
        adaptive_damping: bool = False,
        down_damping_ratio: float = 9.0,
        up_damping_ratio: float = 11.0,
        damping_accept: float = 0.1,
        ellipsoidal_damping: bool = False,
        **kwargs,
    ) -> Optional[torch.Tensor]:
        # "err" tensors passed as input refer to the squared norm of the
        # error vector, as returned by self.objective.error_metric()
        if adaptive_damping:
            return self._check_accept(
                delta,
                new_err,
                previous_err,
                damping_accept,
                down_damping_ratio,
                up_damping_ratio,
                ellipsoidal_damping,
            )
        else:
            return None

seems adaptive damping doesn't work

[EXing]

    optimizer = th.LevenbergMarquardt(
        problem.objective,
        linear_solver_cls=th.LUDenseSolver,
        linearization_cls=th.DenseLinearization,
        step_size=1e-2,
        max_iterations=50,
    )
    theseus_layer = th.TheseusLayer(optimizer)
    with torch.no_grad():
        with global_map.keyframes_mutex:
            print("Start theseus iteration:")
            theseus_outputs, info = theseus_layer.forward(optimizer_kwargs={
                "verbose": global_map.cfg.inner_optim.verbose,
                "adaptive_damping": True,
            })
            print("End theseus iteration.")

This works, but very slow with many iterations

[luisenp]

Hi @EXing do you have a short repro script so I can take a closer look?

[EXing]

Sorry, it's hard to reproduce with a short script.
The project contains multiple files and needs to download datasets.
I can email you a project zip file.