Details
-
Type:
Bug
-
Status: To Do
-
Resolution: Unresolved
-
Fix Version/s: None
-
Component/s: meas_astrom
-
Labels:None
-
Team:Data Release Production
-
Urgent?:No
Description
There are occasional silent failures of the astrometry task in processCcd, that Dominique Boutigny has long complained about on Slack: e.g., https://lsstc.slack.com/archives/C978LTJGN/p1526484056000588 but that never got ticketed.
The gist of the problem is that sometimes the astrometry fails badly (in particular in many DESC DC2 u-band ccds, but there are probably other examples) but the fitter thinks everything is fine. This has led to Dominique Boutigny's script to check for these failures: https://github.com/LSSTDESC/ImageProcessingPipelines/blob/master/python/util/checkCcdAstrometry.py (see DM-16301 for a ticket to add the logic from this script to processCcd).
Recent investigation (see, e.g., https://lsstc.slack.com/archives/C978LTJGN/p1543335472430300 and following) makes it seem like the culprit for most (all?) of these silent failures is that the fitter is trying to fit more SIP terms than there are constraints from the stars. When these stars are in only one section of the CCD, this leads to overfitting for the matched stars and wild extrapolations for other stars, leading to internal statistics from the fitter looking "good" but the overall fit being bad.
(In general, if the default order of the SIP fit is reduced to 2 this seems to greatly increase the success rate on simulated u-band images).
Specifically, for this ticket, I propose having the fitter raise an exception of there are not enough matched stars for the SIP order that was requested. This would turn silent failures into loud failures, as well as alerting the user of where things were going wrong.
Attachments
Issue Links
Activity
John Swinbank While I appreciate the updates to the ticket title, which is more accurate for the actionable part of the ticket, I wanted to make sure that this fix will not necessarily fix all the silent failures. I believe we can just file a new ticket if we find more silent failures after this fix is put in, but we want to make sure that ticket link to this one should it be necessary.
At the same time, is it appropriate on this ticket to change the default order for the astrometry fitting in processCcd?
John Swinbank While I appreciate the updates to the ticket title, which is more accurate for the actionable part of the ticket, I wanted to make sure that this fix will not necessarily fix all the silent failures. I believe we can just file a new ticket if we find more silent failures after this fix is put in, but we want to make sure that ticket link to this one should it be necessary
Understood. As you indicate, the aim here is to acknowledge that tickets are actionable. If you like, we could file another ticket now to audit the codebase and check that there's no possibility of an silent failure, or we could ticket them on a case-by-case basis as we come to them.
At the same time, is it appropriate on this ticket to change the default order for the astrometry fitting in processCcd?
.
John Swinbank, We could definitely fold this into DM-16301, but we promised HSC that LSST would do this, not the other way around.
Everyone vote for this and DM-16301 if you want DRP to do this as part of pair programming this week. Otherwise, it's scheduled for the Astrometry QA sprint in Feb.
@Eli Rykoff: Dan Taranu, Morgan Schmitz and I have been looking into this ticket as part of a pair coding exercise. We've added additional loggers that should ensure that this issue no longer leads to a silent failure. Specifically, if the number of matched sources (numMatches) is less than 10 we raise a TaskError. We also log the number of rejected matches, for reference. Our only outstanding question is how to choose an accurate value for numMatches? The original ticket recommends that the minimum number of matches should vary as a function of the SIP order. We'd like to know if 10 is an acceptable value for us to use here, or if there is a better way to generate this number.
FYI, we've been testing this by running this command:
$ processCcd.py /datasets/DC2/repoRun2.2i --output ../test --id expId=466758 --clobber-config --no-versions
Are your changes specific to the SIP fitter, or generic to all fitters? The Affine fitter should do ok on fewer sources than the SIP fitter.
Whatever that number is, it should be made configurable.
We only added it to FitTanSipWcsTask, but we could to the same for the affine fitter too. Adding a config parameter seems reasonable. Like Lee, I skimmed through the original discussion on slack about the required number of free parameters for a given order and didn't see a conclusive answer. So, I'm not sure what to set the default to or whether it needs to be validated (i.e. raise or warn if the user sets the minimum matches to zero, or less than say ten stars for the default 2nd order).
On a related note: Dominique Boutigny if you have examples of problematic fits that we can reproduce on lsst-dev, that would be very helpful.
On closer inspection, it appears that CreateWcsWithSip.cc (line 130) already implements SIP order checking against matched star numbers (post-rejection), and so the rationale for this ticket may have been removed. Specifically, it checks that the number of matched stars is greater than order+1, which we believe to be the correct logic in this case.
To wrap up this ticket, I propose that we leave in the python-side info logger code we've added which provides the number of rejected stars if greater than zero, as this information wasn't immediately obvious beforehand.
My guess is that that cut is the bare minimum for the code not to crash, but in practice is insufficient to ensure the fit actually makes sense.
I agree - the trouble is that this ticket overlaps with DM-16301, which seemingly better addresses that issue. For the purposes of this ticket (having the fitter raise an exception if there are not enough matched stars for the SIP order that was requested), it looks like this is already being done within CreateWcsWithSip.cc, so we're not sure what else can/should be added here.
Following the discussion on DM-16301, I'm also pausing this ticket until further 'successfully failing' data IDs are available for testing.
Yusra AlSayyad — it'd be nice to get to this one, and
DM-16301, soon. I think you told me the latter was going to be handled by the HSC team — is that right? Any chance of folding this into it?Otherwise, I can put this on the backlog for AP and try to prioritise it.