DM-34531: Make piff PSF model size configurable

[parejkoj]

No description provided.

[arunkannawadi]

LGTM! You might want to wait for Josh's comment just in case.

[arunkannawadi]

Should we have a check= option to ensure it's always odd?

[jmeyers314]

Doesn't this duplicate the base config: https://github.com/lsst/meas_algorithms/blob/main/python/lsst/meas/algorithms/psfDeterminer.py#L37-L51 ?

[parejkoj]

Thanks @jmeyers314 : I hadn't noticed the kernelSize in the parent class. It looks like it might be used rather differently there, though, with the min/max values being related to some calculated kernel size, not the one in the config? I wish those were documented better.

[arunkannawadi]

Josh and I briefly looked into this yesterday, and it looks like the docstring for kernelSize in the base class is incorrect. It reads like it's a measure of size in units of measured sizes, but it really is just the number of pixels along a side to draw a PSF image. It's therefore consistent with the use here.

As for using a RangeField, there's no algorithmic reason to restrict it between 11 and 35, and would rather let that be configurable as well. One thing we could do is to throw an error in validate method if kernelSize exceeds kernelSizeMin and kernelSizeMax. Currently, it is being clipped and I'd rather throw an error instead doing something else silently.

[parejkoj]

Except that, if I'm reading the pcaPsf and psfexPsf classes correctly, kernelSizeMin and kernelSizeMax are used to check whether the kernel size, after scaling by the stellar quadrupole moment, is within some range. So, I don't think those are even relevant to piff at all, since it doesn't do any such scaling. psfexPsf even redefines config.kernelSize with a new docstring an a default of 81.

So, I think we want to revert most of these changes to piff here, but maybe add some clarifying comments to config.setDefaults about how kernelSize is used, and probably drop the usage of min/max entirely?

[jmeyers314]

Except that, if I'm reading the pcaPsf and psfexPsf classes correctly, kernelSizeMin and kernelSizeMax are used to check whether the kernel size, after scaling by the stellar quadrupole moment, is within some range.

I think you're right in principle. In practice, the kernelSize we've been using for HSC at least is 81!

grep kernelSize /repo/main/HSC/runs/RC2/w_2022_08/DM-33741/20220217T181232Z/characterizeImage_config/*.py

gives

... snip ...
config.measurePsf.psfDeterminer['psfex'].kernelSize=81
... snip ...

and is outside the nominal min/max. I don't know about DC2 or other datasets. I suspect we may not have been using the quadrupole scaled stuff in a while, which is why I didn't think to implement that for piff.

[yalsayyad]

Hi! Was directed here by @parejkoj re my question for what default configs to choose.

@jmeyers314 So what does it mean when you say that the psfDeterminer's kernelSize for HSC is 81, when the actual output PSF model kernels are 41 pixels (e.g. the ones you'd get by...

calexp =  butler.get('calexp', visit=30504, detector=54)
calexp.getPsf().computeKernelImage(calexp.getBBox().getCenter()).getBBox()
>>> Box2I(corner=Point2I(-20, -20), dimensions=Extent2I(41, 41))

It says 21 pixels for the post-PIFF butler = Butler('/repo/main', collections=['HSC/runs/RC2/w_2022_16/DM-34451'])

[arunkannawadi]

what does it mean when you say that the psfDeterminer's kernelSize for HSC is 81, when the actual output PSF model kernels are 41 pixels

So for psfex, the kernelSize is set to 81 and samplingSize to 0.5. Since 81*0.5 ~ 41, it takes a 41x41 image around the stars to determine the PSF and generates an oversampled 81x81 image when asked to draw a PSF model.

[arunkannawadi]

Closing this PR without merging, since this is no longer relevant. This has been taken care in DM-36071 and this ticket has been marked as Done.