fixed cost evaluation has floating point issues about ceres-solver HOT 6 CLOSED

Comment #1 originally posted by sandwichmaker on 2012-07-13T15:11:35.000Z:

While trying to familiarize myself with ceres's internals a bit more (I am tempted to try to add a few things), I thought this would be an easy and instructive exercise (although probably as irrelevant as it gets). I have attached the result.

from ceres-solver.

Comment #2 originally posted by sandwichmaker on 2012-07-13T15:15:13.000Z:

Indentation was wrong in a few places.

from ceres-solver.

Comment #3 originally posted by sandwichmaker on 2012-07-13T15:16:49.000Z:

Hey Markus,
This is a great way to start. Keir and I will take a closer look at the patch over the weekend.
Thanks and we would love to know what ideas you have about adding things to Ceres.
Keir and I usually argue about it over a design doc or a rough patch.

Sameer

from ceres-solver.

Comment #4 originally posted by sandwichmaker on 2012-07-13T16:08:13.000Z:

In particular, I was wondering how difficult it would be to add a hybrid Quasi-Newton solver. As the structure (residual and parameter blocks) of the problem is known anyway, it should be easy enough to estimate the Hessian of the i-th residual r_i not as B_i = J_i^T J_i, but B_i = J_i^T J_i + A_i where A_i is computed with a suitable Quasi-Newton update (cf. e.g. Nocedal & Wright). I have used Toint's LANCELOT package before, but firstly (as far as I could see) it completely ignored the J^T J part, and secondly it was far less convenient to use (no auto-diff, Fortran with no C++ frontend, ...)

However, I have no idea how much time I'd be able to invest in this.

from ceres-solver.

Comment #5 originally posted by sandwichmaker on 2012-07-13T17:55:36.000Z:

Re: quasi-newton: Sameer and myself have had similar ideas. In case you noticed, I recently added support for evaluating the gradient in addition to the jacobian. I'm planning to implement nonlinear conjugate gradient, which is different than what you're suggesting.

We'd love to have you work on sparse QN. Instead of writing a big patch, let's have a design discussion on the mailing list.

Having quasi-newton would help with large residual problems.

from ceres-solver.

Comment #6 originally posted by sandwichmaker on 2012-07-18T16:41:08.000Z:

https://ceres-solver-review.googlesource.com/1260 fixes it.
The change is pushed to code.google.com.

Thanks Markus.

from ceres-solver.