Infinities and NaNs in quadratic_prog when c=0 about jaxopt HOT 6 CLOSED

Great that the problem disappeared. Not sure what fixed it, maybe ac1bdcd. See https://github.com/google/jaxopt/commits/main/jaxopt/_src/quadratic_prog.py for the history of this file.

from jaxopt.

In case it helps, for my research code I've manually implemented the same call that QP would've done except using solve_normal_cg and it does solve it without diverging for c=0. It also gives a solution that is similar to that of gmres when setting c small but non-zero. For gmres, letting c go to zero makes it eventually diverge.

from jaxopt.

Hi Ferran

That's weird, I tried to reproduce your bug and did not succeed: https://colab.research.google.com/drive/1-IS1MIkkXfVuON5IhAT2gxt2pw-5IVz8?usp=sharing

Can you give more details on the versions of Python/Jax/Jaxopt you are using ? Are you working on CPU or GPU ? In the default float32 of Jax or float64 ? Is it wrapped in a jitted block or in "eager" mode ?

If you have a notebook in which you can consistently reproduce the bug that would be great.

Is it diverging for primal/dual computation in run(), or during implicit differentiation ?

from jaxopt.

My bad, it was because I had the jaxopt-0.0.1 version; must have been fixed since then!
Just for completeness, it was during the primal/dual computation. Changing the inner solver to solve_normal_cg worked. With jaxopt-0.1.1 gmres (QP's vanilla solver) gives the same result as solve_normal_cg.

Thanks!

from jaxopt.

@mblondel @Algue-Rythme during my research I just ran into more NaNs on the default QP solver, so it may not have been fully solved in 0.1.1. It happens roughly for ~5% of the examples and I haven't found any pattern on when. On the other 95% it matches (modulo small precision errors) the result of using solve_normal_cg as the inner solver.

The failure case is quite entangled with my research code so I can't send it atm. If I find a failure pattern I'll design a minimal example and send it to you.

from jaxopt.

My guess is that the issue is in gmres itself, as our code doesn't seem to contain operations that could result in NaNs.

In #98, @Algue-Rythme is working on a new class EqualityConstrainedQP specific for this case and it will expose the solver, which will allow you to switch to solve_normal_cg.

from jaxopt.