search

facebookresearch / theseus

A library for differentiable nonlinear optimization
MIT License
1.71k stars 121 forks source link

Efficient line search in optimizer #153

Open mhmukadam opened 2 years ago

mhmukadam commented 2 years ago

Current implementation is not friendly for the development of line search and will need some refactoring.

tvercaut commented 1 month ago

I found the lack of line search to be a real bottleneck for the applications we have been looking at so far. For those where differentiability of the solution is not critical, we can rely on scipy.optimize.least_squares and exploit autograd from pytorch throughtorch.autograd.functional.jacobian however this also means having to go back to the CPU which may slow things down depending on the problem size.

Are there any plans to introduce line search to bridge the gap in convergence reliability with scipy, ceres and the like?

luisenp commented 1 month ago

Hi @tvercaut. This feature is a the top of our wish list, but, unfortunately, we are extremely time constrained with other projects at the moment and don't have time for active development on Theseus. We don't have a concrete time line for this at the moment, really sorry about this.