Mosek gp solver is currently not exposed in python about gpkit HOT 5 CLOSED

The email transcript does not appear to be attached! (edit: now it is!)

Should appliedopt also be able to use the CVXopt GP solver?

from gpkit.

I'm prioritizing the exposure of mosek over cvxopt in hopes that we can "fail fast" on using python/numpy, in the event we can't get mosek working from Python. Cvxopt is probably a natural thing to eventually expose, but it comes with a ~30x performance hit (edit: performance hit does not actually appear to be this bad) on the simple problems we've tried. I want to verify we can call mosek right away to ensure we are setting ourselves up to solve big problems.

from gpkit.

I've got something working (besides the command line interface, that was silly but educational) but it seems that, since that they didn't build expopt as a dynamic library anywhere, it'll have to be built by us!

from gpkit.

Annoyingly, their expopt code leaks two debug statements straight to printf, instead of to the task's defined stream . :(

I got it working in the existing framework! But it's only twice as fast as CVXopt on my old computer?!
This even though the profiler says that it's spending >90% of its time running their C code...maybe:

• @whoburg has a computer is much quicker than mine, or
• that simple aircraft example wasn't the example we anticipated a 30x speedup for, or
• Windows is terrible, or
• their matlab implementation is pulling some additional tricks (e.g., multithreading).

from gpkit.

Re: leaking debug statements: once we've got something fairly stable working, I'd be willing to get in touch with them and see if they want to pull in your improvements. They are pretty responsive and might welcome it. Just need to make sure they credit @bqpd .

from gpkit.