Verification Datasets Meeting 2015-12-23 Minutes
Attendees: David, Angelo, Simon, RHL, Gruendl, Russell
- diagnostic plot on # of failures vs. number of sources per ccd
- lots of failures everywhere, but failures quickly rises above ~2000 sources per ccd
- RHL mentioned that there are diagnostic plots, but Simon thinks it might be hard if there are calibration errors
- Angelo will make an ipython notebook that shows how he did things
- RHL suggests turninng on “–debug” and setting debug.py to turn on the debug plots
documentation on how to use the debug option
go to “AstrometryTask” and “Debug Variables”
- debug-level bugging, --loglevel debug, can set component names, trace level with a negative integer
- there is ds9 debug ported from HSC side
- trace gives you more verbosity levels of information, but this might not give more information than using debug
- work is being done on developing multiple levels of debug or trace, RHL will post a community post on this to sort this out
- to show just the astrometry config parameters and documentation
processCcdDecam.py . --show config="*calibrate.astrometry*"
- do the icSrc files get written if the astrometry fails? no
- Angelo will send command to produce the failure and RHL will look at it
- not finding the faintest sources for the ones that failed, check detection config values
- what threshold is that? global sigma, local sigma
config.detection.thresholdType='stdev' global value
- should use “pixel_stdev” which will use a per pixel variance value
- OpenBlas performance issue
- was running 40 processCcdDecam jobs on new linux machine and got very bad performance
- JMatt and I tracked it down, it was an OpenBlas problem (dependency of numpy that miniconda installs)
- OpenBlas uses all possible cores by default, so all processes were trying to use all cores and so 5120 threads were created and all the time was spent swapping
- see ticket https://jira.lsstcorp.org/browse/DM-4714
- the consensus was that the workflow system should set the appropriate global environmental variable (I’ve already implemented this in my IDL workflow) and that this problem needs to be documented.
- JMatt will make a post on community over the holidays on this
- John Swinbank fixed the LTV1/2 error in afw (thanks!), need to rerun processCcdDecam on all images
- discovered that showVisitSkyMap doesn’t actually print out the information that I need and does a crude overlap checking of visits and patches. I’m making a modified version called printVisitSkyMap that will print out visit, ccd, tract and patch for all ccd+patch overlaps.
- giving talk at AAS (at NSF pavilion) on verification work, thanks to everyone that sent figures!
Colin (not on but sent in update):
- Turning on dipole measurement, including sorting out the numerous warning messages that result, and understanding the output catalogs.
- Automating the pairing of exposures to difference. This might mean generating “templates” that the stack recognizes.
Francisco/Yusra: not on
Dominique: not on
Michael Wood-Vasey: not on
No call until after the AAS