I currently have an important code base that is based on the Ceres and Eigen optimization library. I have been told that I must do it fast enough to run it in real time, where it is currently running at 1Hz, and I was hired for that purpose. You are using the Ceres autograd functions.
I was considering a raw CUDA port, but I am also considering putting resources and months of work to port this code in Pytorch, to take advantage of its GPU Tensor operations and possibly also its own autograd features.
I'm curious to know if I'm getting any advantage for this. If this turns out to be a failure, it could cost me a lot in my company