Uploaded image for project: 'Data Management'
  1. Data Management
  2. DM-14172

Optimize cModel config parameters



    • Story Points:
    • Epic Link:
    • Sprint:
      DRP S18-5, DRP S18-6, DRP F18-1
    • Team:
      Data Release Production


      I ran multiBandDriver.py on HSC-R 9813 3,4 to determine if any cModel configuration parameters can/should be changed to improve the running time and/or best-fit model parameters. My conclusions for the tested parameters are:

          initial.nComponents (default 3): For 2, initial fits run in 0.67x the time, but the exp/dev fits run more slowly, so the total fit time is 0.94x. Initial fits are slightly worse on average (with large scatter) and a sizeable fraction of the dev/exp fits are significantly worse (deltalnP < -10).
              For nComponents=6, the initial fitting time is more than twice as long, with modest increases in the exp/dev lnP (<1 for most galaxies, with scatter).
              Recommendation: leave it at 3 or lower to 2.
          initial.optimizer.gradientThreshold (default 1e-2): 1e-3 runs very slightly faster than default and very slightly improves fits. 1e-1 is slightly slower and also results in slightly worse initial/final fits.
              Recommendation: lower to 1e-3.
          initial.optimizer.maxInnerIterations (default 20): Zero impact on fits, running time for 8 is virtually identical; 50 is 1% slower.
              Recommendation: leave at 20.
          initial.optimizer.noSR1Term (default False): True radically changes fitting times by factors of (0.33, 4.37, 4.98)x for (initial, exp, dev) i.e. the initial fit is faster but the subsequent exp/dev are much slower.
              Confusingly, the initial fits are better on average, whereas the final dev/exp are worse - this may warrant follow-up. Many of the very slow fits have faint PSF mags and may be pathological cases.
              Recommendation: leave as False.
          region.nFitRadiiMax (default 3): Surprisingly, nfitrad=2 initial fits are 0.5% slower than the default, whereas nfitrad=5 is 2% faster. However, the total fitting time scales by a factor of (0.86, 1.38) for the smaller/larger size.
              Smaller fitting regions naturally have larger likelihoods, but I can't make a meaningful comparison between likelihoods - presumably the smaller region fits better in the common are and worse outside.

              Recommendation: leave as is pending further investigation.

      One common feature is that there is a non-negligible fraction of objects (usually with faint mag_psf) with very small differences in initial fit likelihood but large ones in the dev/exp fit, which should not occur frequently.

      The quoted figures can be generated by the notebook created in DM-14118 (lsst-dev:/home/dtaranu/src/mine/taranu_lsst/cModelConfigs.ipynb).


          Issue Links



              dtaranu Dan Taranu
              dtaranu Dan Taranu
              Jim Bosch
              Dan Taranu, Jim Bosch
              0 Vote for this issue
              2 Start watching this issue


                Start date:
                End date:


                  No builds found.