IPP Software Navigation Tools IPP Links Communication Pan-STARRS Links
wiki:ppSub_testing_201201

Version 59 (modified by Mark Huber, 14 years ago) ( diff )

--

Summary of ppSub_testing_201201 ---IN PROGRESS PLEASE DO NOT EDIT---

  1. Summary of ppSub_testing_201201 ---IN PROGRESS PLEASE DO NOT EDIT---
    1. Background
  2. Setup
  3. General/Baseline "start" tests
  4. Parameter Space
    1. Kernel parameters
  5. PSF Matching/Convolution
    1. Poor PSF-matching?
    2. No-convolution?
  6. Convolution direction choice
  7. Test Sets
    1. MD04 primary samples
    2. Simulated Samples
    3. 3PI sample
  8. Monitoring
    1. Nightly Tests of Diffims
    2. HOTPANT Diffim Comparisons
    3. diffim stats and flags

The primary purpose has been expanded and redirected. While it appeared that the fixed convolution direction may have been the likely cause of an overabundance of false detections from residuals of bright sources in the field, issues also in the convolved stack with a resulting larger than selected target PSF has illustrated there may also be issues with the PSF matching/convolution itself. So the scope of testing has been broadened/focused to the PSF matching/convolution in ppSub (tests with ppStack will need to follow)

  • is ppSub doing what we want
  • is ppSub doing what we think
  • has something changed along the way (in the data or code)

General tasks (no particular order and by no means complete, concise, clear or quantitative):

  • need to establish a baseline to verify against any past/future changes in code and with reprocessing including a test set of real+simulated images that is readily inter-comparable (avoid having to pull values from logfiles etc)
    • IPP/ppSub: current ops tag ipp-20111222, branches/meh_branches/ppsub_test [available]
    • photpipe/HOTPANTS: V3 configuration locally configured [finally running again on local desktop -- auto choice in progress, forced opposite to setup after. HOTPANTS compiled on all machines]
    • datasets:
      • MD04: for S-S and S-refstk (may want W-refstk later) [stacks and refstack selected over time, IQ and with know sources finished and "bundled"]
      • sim: for all cases W-W, W-S, S-S, S-refstk [in progress, mostly created with past testing with ppStack -- hold with just these]
      • 3PI/other: for W-W and W-S cases [still need to select images to use -- delay for now]
  • wrap up and re-test the convolution direction choice in ppSub (see monitor)
    • does it still work as it did when tested a few months ago with MD10.V2? [to test for the baseline sample MD04.V3 S-refstk and S-S, sim]
    • does it affect the dual convolution [need 3PI/other sample still, need forced photometry setup]
    • if it chooses to convolve the input, is it poorer (see recent monitoring) with greater false detections, larger/messy residuals
  • verify the PSF matching, convolution and diffims - test as expected, self-comparison, cross-comparison (need add a translation map for different outputs)
    • PSF shape - is it equivalent to the target before convolution and subtraction (is there a scaling/normalization issue?)
    • appears solution fairly insensitive to input kernel params --
    • return to stamp choices - need a check for "poor" not being rejected, are masked values set to 0 under stars in background subtracted image reasonable or cause discontinuity problems?
    • forced photometry measurements on sky and selected sources (stats and residuals) --
    • the "no-conv" case - what is limiting ppSub -- have many cases of "close" inputs in test sets, what are the resulting kernels and output convolution params
  • tracking sources of "extra" false diffim detections
    • obviously more use of flags2 and DIFF_* diffim stats -- need to map conditions in more detail (see tuning) as seem to miss flagging the lower level peak/blister residuals (below)
    • lower level peak/blister residuals -- seem to come and go not always aligned with FWHM (see monitoring and MD04 test sample) and not found often in photpipe (to be tested with test set however)
  • monitor new data/reprocessing/tag changes
    • initially new data as MD04 is ongoing and IQ is improved -- convolution direction tests, false detection tests
    • obviously test changes with ppsub_test
  • parameter space, tuning, adding
    • kernels (current ppSub using equivalent HOTPANTS, pandromeda pipeline values) and proper scaling -- currently ppSub actually seems mostly insensitive to the kernel params
    • DIFF_xRATIO diffim stats (photpipe vs MT)
    • values to be added to gpc1 DB and/or the FITS headers (for checking, monitoring)

Background

Samples of previous information/work/comparisons

Setup

Operations tag for testing base diffims is ipp-20111220 (since local tests could/should fix at some r since the tag has revisions added to it? )

Branch for originally testing the choice of convolution direction but now full development testing is also updated to r33098 (from 1/11/2012) -- ​http://svn.pan-starrs.ifa.hawaii.edu/trac/ipp/browser/branches/meh_branches/ppsub_test

Summary of changes so far:

  • for convolution direction choice
    • normalization updated each direction -- need to look if in path for ppStack fixed, seem to recall modifies even fixed (may be for both)
  • for PSF matching
    • kernel scaling based on Kron radius instead of
  • general, help for MOPS using DIFF_R_P for cuts -- ppSub/src/ppSubFlagNeighbors.c DIFF_R_P/M chooses closest match based on detection peak flux s/n, change to 1/psfmagerr used for DIFF_SN_P/M.
    • to be tested how responds to if no input catalogs provided -- no catalogs produce an error before the change 
      Error in subtraction:
 -> subScaleKernel (ppSubMatchPSFs.c:127): Problem in data values
     Unable to scale kernel, since no sources were provided.
 -> ppSubMatchPSFs (ppSubMatchPSFs.c:379): Problem in data values
     Unable to scale kernel parameters
 -> ppSubLoop (ppSubLoop.c:113): Problem in data values
     Unable to match PSFs.
    • only the no-conv case allows no input catalogs, so need to trace what detection peak is used for that -- running NOCONVOLVE TRUE then faults with Assertion failed in function ppSubMakePSF at ppSubMakePSF.c:61. Error stack: Abort. -- so don't need to worry about case of no input sources and can make change.
  • gpc1 DB columns:
    • FWHM/IQ would be helpful to have in stackSumSkyfile
  • FITS keywords:
    • kernel sum useful in ppSub?

Photpipe/HOTPANTS setup on desktop to start, then may move to the dev cluster (HOTPANTS also build on dev and ops clusters).

General/Baseline "start" tests

Parameter Space

Kernel parameters

HOTPANTS suggested:

  • need to also resolve size/sigma conversion between ppSub and HOTPANTS

M31 group:

  • standard deviations of 9, 6, 3 and 0.1 pixel and a polynomial order of 2, 4, 6 and 0
  • independent but similar algorithm described in Kerins et al. 2010 ​http://arxiv.org/abs/1004.2166 for regions with complex backgrounds
  • look at difference in photometric scaling method in more detail

PSF Matching/Convolution

Poor PSF-matching?

Not sure yet what to classify this as other than the PSF not being matched well. Suspect similar to issue with convolved stacks in ppStack.

No-convolution?

Reports from other surveys that diffims are poorer with images having similar PSFs, convolution isn't well matched? Is this for dual only?

Convolution direction choice

We want to avoid large deconvolution in PSF-matching image subtraction. As the image quality of the images continues to improve, always convolving the reference stack will not produce the best difference image and convolving the input image is desired. The quality of the difference image is evaluated in both directions and the better is chosen for the output (need to add specific condition test).

MD04 sample:

Not all seem to be really due to convolving in the wrong direction and need to identify and separate the cases/causes if can. Running samples forced in both direction useful to see.

CfA photpipe/HOTPANTs direction: using single template/reference stack so locally get keyword CONVOL00 from header except slight issue with amps. All are convolving the TEMPLATE only so also suggest issue seems not a convolution direction but convolution issue itself.

  • discovered the current photpipe setup is now set to for convolving the template only.. rather then testing both.. argh..

Sample diffims with direction forced and detections marked (note both stack-refstack and stack-stack cases). The stack-stack case is the really the only way to test the convolution direction choice now unless go back to the MD04.V2 sample. (easiest to use event stamps from pages, but may want image photpipe diffim..)

  • ref - ref
  • stk - ref vs ref - stk (both directions)
  • stk - stk (both directions)
obsdate/stack_id (stk-ref) FWHM_MAJ (pixel) IQ_FW1 (pixel) ppSub (auto choice) photpipe/HOTPANTS (auto choice) comment
refstack.20111012 / 450933 5.206454 4.08811 -- --
20100110 / 444020 4.878624 3.891729 2 i
20100121 / 444079 5.843588 4.574009 1 t
20110129 / 444251 4.424473 3.378561 2 i
20110201 / 644876 5.746622 4.748873 1 t
20110312 / 446658 4.958342 3.74775 2 i
20110315 / 645114 4.270276 3.232871 1 i
20110423 / 446715 4.603964 3.510957 2 i
20110429 / 645474 7.178107 6.027955 1 t
20120115 / 622752 5.805826 4.622056 1 t
20120121 / 646801 6.346761 5.359003 1 t
20120126 676389 4.443175 3.475093 1 i

Example cutouts -

event (auto direction) TARGET2 (convolved ref) / TARGET1 (convolved stk) photpipe automatic (direction) comment
X (TARGET2) No image "test.png" attached to ppSub_testing_201201 No image "test.png" attached to ppSub_testing_201201 (TEMPLATE) conv refstk better,

Photometry/detection comparisons:

Simulated sample:

Test Sets

Need tests to verify changes will not break diffims for other surveys (MOPS etc) but do not need to worry about magic now. Not for testing if they will choose the right direction (only if using AUTO), but changes involved part of the normalization and will need to verify doesn't change flux/variance values:

  • MD
  • simulated image set WW,WS,SS, SSref -- since SSref for MD04 seem to prefer convolving the deeper refstack still, will need to generate
  • warp-warp diffim: MD and 3PI TTI
  • warp-stack diffim: MD and LAP stack

MD04 primary samples

Initially selected for if choosing the proper direction in stack-stack diffims now using MD04. Was previously done using MD10, but we want to continue to build up a test set for testing other parts of the IPP and MD04 is a favorite with extensive overlap from other surveys.

Start with just 2 skycells having example events from last year for photometry comparisons as well -

skycell.055 (previous tessellations 077) -- near center -

skycell.058 (previous tessellations 073) -- edge -

QUB cross-match page: ​http://star.pst.qub.ac.uk/ps1/psdb/crossmatchipptocfa/ (add when time.. difficult to search the new crossmatch page..)

Difficult to use past examples of wrong convolution direction since refstacks remade, so select sample to just test current choices. The monitoring part should take care of of this in more recent cases as MD04 actively being observed for next couple months.

QUB examples of wrong convolution choice:

Sample nightly stacks for both skycells with same obsdate to look at variation of PSF center and edge (0.25 "/pixel), limited to r-band for now and a sample size that can be manually rerun with some ease.

still need to find a night with very similar IQ

obsdate stack_id FWHM_MAJ (pixel) IQ_FW1 (pixel)
skycell.055
refstack.20111012 450933 5.206454 (+-0.6-0.8) not the 4.5 on the MD.GR0 page.. 4.08811 (+-0.3-0.4)
20100110 444020 4.878624 3.891729
20100121 444079 5.843588 4.574009
20110129 444251 4.424473 3.378561
20110201 644876 5.746622 4.748873
20110312 446658 4.958342 3.74775
20110315 645114 4.270276 3.232871
20110423 446715 4.603964 3.510957
20110429 645474 7.178107 6.027955
20120115 622752 5.805826 4.622056
20120121 646801 6.346761 5.359003
(swap 20120126 for 20120115?) 676389 4.443175 3.475093
skycell.058
refstack.20111012 450936 4.616929 3.926292
20100110 444023 4.777016 3.687181
20100121 444082 5.567342 4.243489
20110129 444254 3.976701 3.100284
20110201 644879 5.023516 4.421264
20110312 446661 5.114515 3.701454
20110315 645117 4.58284 3.376984
20110423 446718 4.358292 3.138227
20110429 645477 6.088856 5.447231
20120115 622755 5.300241 4.337034
20120121 646804 5.627441 5.058287
(swap 20120126 for 20120115?)

Simulated Samples

Simulated image sample:

filename sim FWHM FWHM_MAJ (pixel) IQ_FW1 (pixel)

3PI sample

Will SAS provide enough variation?

Monitoring

Nightly Tests of Diffims

Is there any conflict with the expected desired direction? Compare to CfA photpipe, detection counts.

  • having IQ/FWHM keywords in stack images or at least the gpc1 DB would be helpful here since are calculated for the PSF matching in the diffim process, doing manually here takes a good amount of time.

Is the convolution not optimal?

HOTPANT Diffim Comparisons

Difficult to automate and may want to force both directions for comparison as well as shallower reference stacks to use. Are the diffims continuing to produce similar results.

  • for select night samples should setup and run locally. still on the todo list.

diffim stats and flags

There are columns of diffim stats and flags to help characterize the diffim and detections

DIFF_ equivalent setting from photpipe:

  • DC_FRATIO_MIN 0.75
  • DC_NRATIO_BAD_MIN 0.7
  • DC_NRATIO_MASK_MIN 0.6
  • DC_NRATIO_ALL_MIN 0.5
  • DC_NGOOD_MIN 3

DIFF_ initial settings from Maria Teresa's tests:

  • DIFF_NPOS >5 nPos (n pix > 3 sigma)
  • DIFF_FRATIO >0.7 fPos / (fPos + fNeg)
  • DIFF_NRATIO_BAD >0.7 nPos / (nPos + nNeg)
  • DIFF_NRATIO_MASK >0.7 nPos / (nPos + nMask)
  • DIFF_NRATIO_ALL >0.5 nPos / (nGood + nMask + nBad)

Grid of expectations for DIFF_ values/cases:

Attachments (26)

Note: See TracWiki for help on using the wiki.