About accelarating the conjugate gradient slover in Manopt.

[grandowife]

Hi, there.

I have a problem on Grassmann Manifold, and I choose CG solver in Manopt to optimize my problem.

It indeed works, however, it also costs a lot of time to converge.

Besides, I see a relatively small decrease in 'cost Val', but a relative reasonable decline in 'grad.norm', which are both shown below.

Now,  would you please tell me :

1) how can I accelerate the process of optimization? Is decreasing the 'minstepsize' value in CG works? 

2) how come the loss of my problem nearly unchanged?

Thank you.

[Nicolas Boumal]

Hello,

To be sure: did you run checkgradient on your problem?

Changing minstepsize is unlikely to have a marked effect.

Did you try running trustregions? (It will approximate the Hessian automatically if you did not provide it).

Best,

Nicolas