The purpose of this issue was really to confirm that the translation of the multi-Gaussian estimation using ngmix/EM was actually producing results which could be used by our framework, so CModel was run after EmPsfApprox using the latter as its psfName.

This test actually discovered that we were not running the EMRunner in ngmix correctly. It needs at least a "unitJacobian" to function correctly. Prior to supplying that, the algorithm seemed to fail unpredictably. This is really work which ended up on 6123, as I did these two simultaneously.

The code to run comparisons is actually in galaxy_shear_experiments, under tickets/DM-6124. I state the results here. The TripleGaussians are 3 component Gaussian Models defined for ShapeletPsfApprox, and are 000 and 000. They were installed just for comparison.

The big differences actually seem to be in comparing nGauss1 vs. SingleGaussian. Please note that EMRunner does not run to as small a tolerance as PsfShapeletApprox, as I observed visually when I ran both plugins on pure Gaussians.

CModel comparison: for nGauss2 vs DoubleGaussian:
e1 differences: 0.00000028 + 0.00000180 for 10 sources
e2 differences: 0.00000025 +- 0.00000091 for 10 sources
CModel comparison: for nGauss1 vs SingleGaussian:
e1 differences: 0.00000573 + 0.00002557 for 10 sources
e2 differences: 0.00000453 +- 0.00000907 for 10 sources
CModel comparison: for nGauss3 vs TripleGaussianInner:
e1 differences: 0.00000048 +- 0.00000156 for 10 sources
e2 differences: 0.00000008 +- 0.00000254 for 10 sources
CModel comparison: for nGauss3 vs TripleGaussianOuter:
e1 differences: 0.00000059 +- 0.00000250 for 10 sources
e2 differences: 0.00000043 + 0.00000161 for 10 sources
CModel comparison: for SingleGaussian vs DoubleGaussian:
e1 differences: 0.00000048 +- 0.00000875 for 10 sources
e2 differences: 0.00000427 + 0.00000606 for 10 sources

Differences in the pixels produced by the different plugins vs. the actual Psf given to them. These were PhoSim Psfs.

SUMMARY using PsfApprox SingleGaussian with 10 exposures
PsfApprox runtime: 0.0081 sec. +- 0 sec.
CModel runtime: 0.0061 sec. +- 0 sec
PSF avg pixel diff: -0.00015557
PSF std pixel dff stdev: 0.00248383

SUMMARY using PsfApprox DoubleGaussian with 10 exposures
PsfApprox runtime: 0.0463 sec. +- 0 sec.
CModel runtime: 0.0099 sec. +- 0 sec
PSF avg pixel diff: -0.00014726
PSF std pixel dff stdev: 0.00249288

SUMMARY using PsfApprox nGauss1 with 10 exposures
PsfApprox runtime: 0.0014 sec. +- 0 sec.
CModel runtime: 0.0130 sec. +- 0 sec
PSF avg pixel diff: -0.00016635
PSF std pixel dff stdev: 0.00229705

SUMMARY using PsfApprox nGauss2 with 10 exposures
PsfApprox runtime: 0.0191 sec. +- 0 sec.
CModel runtime: 0.0106 sec. +- 0 sec
PSF avg pixel diff: -0.00016176
PSF std pixel dff stdev: 0.00235760

SUMMARY using PsfApprox nGauss3 with 10 exposures
PsfApprox runtime: 0.2348 sec. +- 0 sec.
CModel runtime: 0.0159 sec. +- 0 sec
PSF avg pixel diff: -0.00016079
PSF std pixel dff stdev: 0.00236489

Here is runtime information which came out of these same tests:

SUMMARY using PsfApprox nGauss3 with 10 exposures
PsfApprox runtime: 0.2366 sec. +- 0 sec.
CModel runtime: 0.0164 sec. +- 0 sec
SUMMARY using PsfApprox SingleGaussian with 10 exposures
PsfApprox runtime: 0.0081 sec. +- 0.003272 sec.
CModel runtime: 0.0061 sec. +- 0.001346 sec
SUMMARY using PsfApprox DoubleGaussian with 10 exposures
PsfApprox runtime: 0.0463 sec. +- 0.005585 sec.
CModel runtime: 0.0098 sec. +- 0.002891 sec
SUMMARY using PsfApprox nGauss1 with 10 exposures
PsfApprox runtime: 0.0014 sec. +- 0.000054 sec.
CModel runtime: 0.0130 sec. +- 0.013712 sec
SUMMARY using PsfApprox nGauss2 with 10 exposures
PsfApprox runtime: 0.0188 sec. +- 0.000400 sec.
CModel runtime: 0.0112 sec. +- 0.001392 sec
SUMMARY using PsfApprox nGauss3 with 10 exposures
PsfApprox runtime: 0.2340 sec. +- 0.009426 sec.
CModel runtime: 0.0158 sec. +- 0.002432 sec
SUMMARY using PsfApprox TripleGaussianOuter with 10 exposures
PsfApprox runtime: 0.3727 sec. +- 0.233824 sec.
CModel runtime: 0.0151 sec. +- 0.003921 sec
SUMMARY using PsfApprox nGauss3 with 10 exposures
PsfApprox runtime: 0.2347 sec. +- 0.009666 sec.
CModel runtime: 0.0158 sec. +- 0.002065 sec
SUMMARY using PsfApprox TripleGaussianInner with 10 exposures
PsfApprox runtime: 0.3720 sec. +- 0.230313 sec.
CModel runtime: 0.0149 sec. +- 0.003539 sec

The code which produced these results is in galaxy_shear_experiments. I am not concerned about checking this in, but if you want to see the code, it is on tickets/DM-6123

Sorry. Actually, DM-6124

This is a comparison of the Gaussian fitters using their default settings. This was just a sanity test. But I did find the omission of the Jacobian because of strange problems with EMPsfApprox doing these simple tests.

nGauss1: ngmix EMPsfApprox, nGauss = 1
nGauss2: ngmix EMPsfApprox, nGauss = 2
SingleGaussian: PsfShapeletApprox
DoubleGaussian: PsfShapeletApprox

Models running on a Single Gaussian:
component 0: amp = 1.000000, sigma = 4.000000

Using model nGauss1
DIFFS: (-0.00015829129176376792, 0.0011888552335592665)
PSF: (ixx=15.9990375642, iyy=15.9990375678, ixy=-1.59644452847e-08),
(-5.7088e-13, 7.2717e-14))
[ 0.27003514]

Using model SingleGaussian
DIFFS: (-0.00015541221422079763, 0.0012018456926277991)
PSF: (ixx=15.9999987904, iyy=15.9999994105, ixy=-2.88753774136e-07),
(7.0562e-10, 5.9369e-10))
[ 0.28209478]

Using model nGauss2
DIFFS: (-0.00015828218642075919, 0.0011889808463263511)
PSF: (ixx=15.437239682, iyy=15.4376494832, ixy=-0.000212610070082),
(0.0019808, -0.0022363))
[ 0.15543501]
PSF: (ixx=16.7722790887, iyy=16.7717162863, ixy=0.000298129770636),
(-0.0027039, 0.003053))
[ 0.11461417]

Using model DoubleGaussian
DIFFS: (-0.0001554122121517216, 0.0012018456811670592)
PSF: (ixx=15.9999999845, iyy=16.0000010528, ixy=-2.86920951877e-07),
(1.1837e-09, 9.7758e-10))
[ 0.28209478]
PSF: (ixx=63.9982986901, iyy=63.9982986901, ixy=1.49369523773e-12),
(-1.9056e-17, -2.3531e-17))
[ 1.48330137e-08]

-------------------------------------
-------------------------------------
Models running on a double gaussian
component 0: amp = 1.000000, sigma = 4.000000
component 1: amp = 0.500000, sigma = 8.000000

Using model nGauss1
DIFFS: (-0.00016236985259901842, 0.00093493041720977275)
PSF: (ixx=26.4048038208, iyy=26.4048038208, ixy=-8.42381887074e-13),
(-3.7257e-16, -6.1202e-19))
[ 0.26343557]

Using model SingleGaussian
DIFFS: (-0.00016284465841898458, 0.00095982776760274104)
PSF: (ixx=20.3547972164, iyy=20.2284155693, ixy=0.0631870250436),
(-0.00017337, -9.5825e-05))
[ 0.25573096]

Using model nGauss2
DIFFS: (-0.0001595243252360783, 0.00095397063580048111)
PSF: (ixx=15.9668132914, iyy=15.9668132914, ixy=0.0),
(3.8047e-16, -5.4508e-17))
[ 0.18150405]
PSF: (ixx=63.7248239325, iyy=63.7248239325, ixy=-3.79858632292e-16),
(2.0634e-16, 2.0936e-17))
[ 0.09142004]

Using model DoubleGaussian
DIFFS: (-0.0001572490117598259, 0.0009619403876283553)
PSF: (ixx=16.169975098, iyy=16.1699215778, ixy=-1.24287220881e-05),
(7.8807e-07, 2.8255e-06))
[ 0.19159612]
PSF: (ixx=67.2220132981, iyy=67.2225928636, ixy=0.000126914960634),
(2.2892e-07, 4.0119e-07))
[ 0.09126995]

Note that nGauss2 finds a two Gaussian fit to a single gaussian which splits the flux between two gaussians of nearly equal sigma and flux.

Looks good. One question: did finding and fixing the Jacobian bug improve the speed of the algorithm at all? No need for lots of new tests; I was just curious whether the slow performance you saw before went away at all.

Yes, I was just sending you email about that. The EmPsfApprox algorithm seems fine now. Note comparison with ShapeletPsfApprox models of similar order.

I think that all the questions for 6124 and 6125 were answered by the timing information I sent you. Yes, I think the plugin is working well now, and is faster. It is about as good as the comparable lower order ShapeletPsfApprox models at reproducing the input Psf. And the CModel + EmPsfApprox pair seems to work comparably.