search

ratt-ru / CubiCal

A fast radio interferometric calibration suite.
GNU General Public License v2.0
18 stars 13 forks source link

jhj weirdness in diag-diag mode #227

Open landmanbester opened 6 years ago

landmanbester commented 6 years ago

Been testing the diagonal solver and found some weirdness. Basically found that cubical was flagging everything even on a high SNR run with simulated data. Turns out if I call the cubical's _solve_gains from outside of cubical (as in here) I get a jhjinv which is about 1e10 smaller than the jhjinv reported in the gain solution tables produced by cubical. Must be related to https://github.com/ratt-ru/CubiCal/issues/224. Not sure if this is only relevant to diag mode but suspect it must be related to cubicals noise whitening

ratt-priv-ci commented 6 years ago

This might explain I struggled with my recent phase calibration (lowsnr) after averaging with oldsplit. I used the diag-diag moes with phase only calibration and kept getting low snr on the slope solver.

On Fri, 19 Oct 2018, 19:23 Landman Bester, notifications@github.com wrote:

Been testing the diagonal solver and found some weirdness. Basically found that cubical was flagging everything even on a high SNR run with simulated data. Turns out if I call the cubical's _solve_gains from outside of cubical (as in here https://github.com/landmanbester/rooibos-calibration/blob/1ba3205f00087a858d7a7a66514dc837fafd69ea/SolutionIntervals/solver_utils.py#L48) I get a jhjinv which is about 1e10 smaller than the jhjinv reported in the gain solution tables produced by cubical. Must be related to #224 https://github.com/ratt-ru/CubiCal/issues/224. Not sure if this is only relevant to diag mode but suspect it must be related to cubicals noise whitening

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/ratt-ru/CubiCal/issues/227, or mute the thread https://github.com/notifications/unsubscribe-auth/ARFXpi0ddJH7P68OWHoPCu3c4GItah6_ks5umgqkgaJpZM4XxG2C .

o-smirnov commented 6 years ago

Yeah I'm pretty sure I see the problem. The error estimates are not re-weighted, so if the mean weights are not 1, the errors aren't scaled right. @IanHeywood you might want to disable flagging on gain errors (set --g-max-prior-error, --g-max-post-error to 1e+99) for now.

JSKenyon commented 4 years ago

@o-smirnov Was this ever fixed?

o-smirnov commented 4 years ago

I don't think so...