Status: Won't Fix
Fix Version/s: None
Turning on doScaleDiffimVariance across the board fixed HiTS. Making sure doBrighterFatter is the same for Gen2 and Gen3 fixed most - but not quite all - of HSC/COSMOS. Both of those things happened in
DM-29344. However, chronograf shows the ratio of DIA to direct sources is still ever-so-slightly different.
The ratio is 4.81 for Gen2 and 4.75 for Gen3 (visit 59160, detector 51)
The ratio is 7.29 for Gen2 and 7.20 for Gen3 (visit 59150, detector 50)
- relates to
DM-29344 Investigate the CI differences between Gen 2 and 3 in COSMOS field
I'm not sure where the CI output repos live. Krzysztof Findeisen may know? So I quickly kicked off my own ap_verify runs for you, using the latest weekly, which should have dropped after I merged
DM-29344. They will be in /project/mrawls/ap_verify_runs/ DM-30055-gen2 and DM-30055-gen3 in the same directory. There are logs in the same directory too, in DM-30055-gen2.out and DM-30055-gen3.out. Both runs should finish shortly.
Jenkins does have scratch space in /project/jenkins/prod/, but it's hard to match up directories to runs, and I, at least, don't have read permission for the actual repos. But if we can't reproduce by running ap_verify anywhere, that would be scary for a different reason...
I just had a look at the results from SFM that Meredith pointed to above. Things are looking pretty identical at that stage for both calexps of visit-ccd: 59150-050 59160-051 (those are the only two I see in /project/mrawls/ap_verify_runs/
DM-30055-gen2/output/01527/HSC-G/corr/) having looked at the usual "all the things" (and the "pretty" qualifier is again just the offset in deblend_peakId when non-zero). Is it possible some "downstream" configs are out of sync?
Thanks Lauren. Those are indeed the only two visit+detector combos in the CI dataset, and what you found jives with what I expected. What do you think Eric Bellm - is this "close enough to identical" or in need of further examination?
Chris Morrison [X] indicated he saw floating-point differences when reading the raws--if he can isolate that let's spawn a new middleware ticket and close this one.
Could this be due to the issue resolved in
When I did the comparison noted above, it included:
- calexp image arrays (image, variance, mask planes)
- calexp photoCalib objects
- calexp PSFs
- calexp WCSs
- every column in the source tables
which, modulo the deblend_peakId caveat, were all identical. As such, I don’t think either of the previous conjectures could be at play.
Gen 2 CI for AP has been turned off, so we are no longer planning to track down this difference.
Can you point me to the new gen2 & gen3 HSC/COSMOS runs?