AP Pipeline QA Meeting, 2021-09-13

4pm Pacific Time

BlueJeans: https://bluejeans.com/426716450

attending:

• Eric Bellm 
• Andrew Connolly 
• Bryce Kalmbach 
• Krzysztof Suberlak 
• Unknown User (cmorrison) 
• Ian Sullivan 
• Colin Slater 
• Ian Sullivan 
• Jeffrey Carlin 
• Kenneth Herner 
• Krzysztof Findeisen 
• Meredith Rawls 
• Spencer Nelson 
• Yusra AlSayyad 

regrets:

Topics for discussion:

• Precursor processing (AP team):
  • Eric Bellm completeness and purity tests of HSC-COSMOS:
    • completeness (fake injection): DM-31340
      • overall lower than expected completeness, and falls off well before the SNR~6 estimated from measured DIASources
      • followup ticket DM-31388 - Getting issue details... STATUS
        • are limiting mags wrong due to detection vs measurement SNR?
          • Chris: use sky sources to get limiting mags
          • plot number counts of direct image sources and look for turnover
            • Eric Bellm double check that matched fakes catalogs are cut appropriately spatially–spatial cuts shouldn't be needed 27 Sep 2021 
              
          • are there fakes being injected on masked regions, or otherwise bad data?
            • → make scatter plots of unmasked fakes
          • Yusra AlSayyad wonders if this is because the template doesn't have enough coverage and doesn’t have enough data to pick the variance plane scale factor right.  Current config is 0.1, maybe needs to be higher.
    • purity (Zooniverse): DM-31341 - Getting issue details... STATUS
      
      • okay but not amazing purity.  flag cuts help but don't give perfect output
        • lots of human-labeled "edge" examples that aren't flagged edge
          • are the masks too small?
          • are these in fact not on the edge?  (maybe interpolated regions or something else)?
            • check xy pixel coords
              • followup: Analysis on DM-31341 shows that our human labels are **better** than the default base_PixelFlags_flag_edge flag
        • lots of "no detection" cutouts
          • Yusra thinks these are bad bg subtraction
            • spatial plots of zero-flux DIASources?
              • followup: Analysis on DM-31341 shows only 1-3 of these per CCD-visit, which doesn't suggest background
          • Chris wonders if this could be those detection footprint outside the detector, which yield nonsense centroids?
            • followup: Analysis on DM-31341 shows that the measured psFlux SNRs are mostly > 5, so the detection centroids are probably fine, and this might be some weird interaction with the PSF model and a few pixels in the variance plane?
          • can't be sky sources as they are excluded from the APDB
          • need to propagate out detection SNR.
            • DM-31811 - Getting issue details... STATUS
  • Kenneth Herner  DC2 5-year depth tract (DRP-mode diffim)
    
    • DM-30003 (needs a Large RSP instance for memory)
    • u/kherner/2.2i/runs/test20210807-tract4431-w32-step5
    • virtually all tracts/patches successfully processed
    • truth table matches
    • 80% purity, 99% completeness vs. truth table
      • this seems better than DESC reports on
      • Yusra likes number counts per square degree as a function of sigma as an absolute metric: completeness/purity work best as comparison on same datasets.
      • Colin: depends if false positives are coming from noise misestimation (area of sky), or objects themselves
• SITCOM team updates (Andrew Connolly  et al.):
  • Krzysztof Suberlak : showing AuxTel processing of donut images–work in progress.  seems to be running slower than expected, debugging.  Aim to use for focus loops
  • Andy: after finishing focus loop closures, will look at WCS, deep learning for donut modeling
• PreOps V&V (Colin Slater ): 
  • mired in scheduling for DP0.2
  • will look at Ken's notebook of DRP-mode processing of DC2 data
• Review CI (https://chronograf-demo.lsst.codes/):
  • no changes!
• Next meetings (all at 4pm Pacific):
  • October 4
• AOB
  • none

Action Items