search

ratt-ru / pfb-imaging

Preconditioned forward/backward clean algorithm
MIT License
7 stars 5 forks source link

Cygnus A #24

Closed landmanbester closed 3 years ago

landmanbester commented 4 years ago

Here is what I get after 10 major cycles (5 reweighting steps)

cyga_before_flagging

Not the best image but, considering that this is with natural weighting, maybe also not the worst. I think some of those artefacts will go away after flagging outliers and rescaling the weights as a function of baseline length. Busy running my flag_outliers script, will report progress here.

Interestingly the l21 reweighting works well for Cygnus A but not on ESO137 and I have no idea why...

o-smirnov commented 4 years ago

Nice! This is A+B+C data?

landmanbester commented 4 years ago

Yep, and D. I actually get a pretty good result after only 2 major cycles but the reweighting steps seem to be necessary to sharpen the image (especially the core) and get rid of some artefacts. Maybe not surprising since its a small fov

landmanbester commented 4 years ago

Are you worried about missing diffuse flux in the lobes?

o-smirnov commented 4 years ago

Not over a PNG rendering. Gonna take a lot of comparing in a FITS viewer before I can answer that question usefully...

landmanbester commented 4 years ago

Hmmm, I am a bit worried. Here is a more saturated image

cyga_before_flagging_saturated

I think it is just swamped by calibration errors at this stage. Let's see what happens after a round of flagging (which made a big difference when I imaged just the D config data). I should also probably set up better diagnostic plots. Plotting whitened residual visibility amplitudes as a function of baseline length should be useful. However, I am not sure if we should look at these per measurement set and spw separately or globally. My current script flags outliers based on how far they are from the global mean of the whitened residual visibility amplitudes. Thoughts?

o-smirnov commented 4 years ago

Are the hotspots subtracted in this data, I forget now?

landmanbester commented 4 years ago

No, that didn't work. I was getting big negative holes where the hotspots are supposed to be. I think clean puts delta components there which are too "sharp" to model with the l21 prior. Maybe we should subtract only a fraction (say 90%) of the hotspots?

audreyrepetti commented 4 years ago

Hi Landman,

I guess adding a calibration step would be more for next stage? :-)

On 9 Oct 2020, at 10:51, Landman Bester notifications@github.com wrote:

No, that didn't work. I was getting big negative holes where the hotspots are supposed to be. I think clean puts delta components there which are too "sharp" to model with the l21 prior. Maybe we should subtract only a fraction (say 90%) of the hotspots?

— 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/pfb-clean/issues/24#issuecomment-706085659, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADLYNP5WQURVQASMMUWVCSTSJ3MJZANCNFSM4SJZ4JMQ.

landmanbester commented 4 years ago

Yes, calibration will come later. I think the next step is to solve simultaneously for an image decomposed into a sum of diffuse flux, point sources and an artefact image. Then we can calibrate with the diffuse flux + point sources (and eventually alternate using a block coordinate pfb algo as you have done in your paper). Having a clean separation between point sources and diffuse flux would be very cool

o-smirnov commented 4 years ago

Ah well then you're definitely up against the artefacts now. Not going to get much more mileage until we get rid of the hotspot artefacts.

Maybe we should subtract only a fraction (say 90%) of the hotspots?

Or add them back in with unity DD-gain? Hmm...

landmanbester commented 4 years ago

I'm not sure I follow. The problem is that the postivity constraint + over-subtracted hotspots (for the l21 prior) means the algo doesn't make much progress. I don't think adding the hotspots back in after the fact will improve things. Is that what you meant?

I think its worth checking how far we can get with flagging, rescaling l2 weights and more l21 reweighting steps first. Maybe that gives an image we can calibrate with. Although DD calibration might be tricky without a component model. I guess we can just make a mask with the hotspots and use those for DD calibration

o-smirnov commented 4 years ago

No, I was thinking of something more Machiavellian -- subtract the hotspots using DD calibration, then add their model visibilities back in without a DD gain...

landmanbester commented 4 years ago

Oh, I see. That mght improve things. Its worth testing if you can find the time. I have continued the imaging run after flagging 5-sigma outliers for now

landmanbester commented 4 years ago

Some progress after phase only selfcal (flagging the divergent tiles was the key here)

cyga_phase_only

but saturating it further shows that there are still some calibration issues

cyga_phase_only_saturated

I can get rid of them by increasing the strength of the L21 regulariser but then the residuals end up with a lot of flux left in them. The current residual is nothing to write home about

cyga_phase-only-residual

I am going to try solving for an additional auxiliary field with a mask where there is still real flux left in the residual i.e.

feature_mask

This is cheating a little bit but I think the resulting image should be good enough for a full amplitude and phase calibration.