search

ACEsuit / ACE.jl

Parameterisation of Equivariant Properties of Particle Systems
65 stars 15 forks source link

Multispecies fitting #6

Closed casv2 closed 4 years ago

casv2 commented 4 years ago

Doing a multi species fit generally results in a large condition number R (e.g. 1e164) for which doing a regular QR decomposition results in a LAPACK error. Performing the fit with RRQR does work.

gabor1 commented 4 years ago

Rather than just powering through it, is it reasonable to see these huge condition numbers? is there not a bug?

-- Gábor

Gábor Csányi Professor of Molecular Modelling Engineering Laboratory, University of Cambridge Pembroke College Cambridge

Pembroke College supports CARA. A Lifeline to Academics at Risk. http://www.cara.ngo/

On 17 Jun 2020, at 16:17, casv2 notifications@github.com wrote:

Doing a multi species fit generally results in a large condition number LsqDB (e.g. 1e164) for which doing a regular QR decomposition results in a LAPACK error. Performing the fit with RRQR does work.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

cortner commented 4 years ago

That’s why I asked Cas to open an issue

cortner commented 4 years ago

@casv2 Does the workaround take care of the condition number issues as well?

cortner commented 4 years ago

We won't close this issue yet though until I've fixed the bug properly. At the moment it is just a workaround!

casv2 commented 4 years ago

Yes condition number seems fine right now!

cortner commented 4 years ago

since the new workaround should not affect performance anymore and since Cas' latest tests seem to confirm this is now all resolved, I'll close this (moving the remaining puzzle to #7) . (fixed as of JuLIP v0.9.6 and SHIPs v0.6.6 )

@casv2 Please feel free to reopen if another issue arises.