Rubin deep coadd simulation code by jchiang87 · Pull Request #521 · LSSTDESC/imSim

jchiang87 · 2026-09-24T04:23:30Z

No description provided.

…oughout

…nsider band-dependence in the internal caches; fix pixel cell center calculation

…are nJy

…ze calculation

rmjarvis · 2026-09-25T13:57:41Z

Is this ready for review? You still have it marked as draft.

jchiang87 · 2026-09-25T14:53:56Z

I was planning to remove the fft stuff before marking it ready, but everything else is ready to be looked at and I expect there will be some iteration anyway, so I'll mark it ready for you to have a look.

… to deep coadd config

rmjarvis

Looks good Jim. Just a few comments and questions.

rmjarvis · 2026-10-15T14:34:21Z

+        params, safe = galsim.config.GetAllParams(config, base, req=req)
+        deep_coadds = galsim.config.GetInputObj('deep_coadds', config, base,
+                                                'RubinDeepCoaddBandpass')
+        return deep_coadds.getBandpass(params['band']), safe


Why is this a special deep_coadds version? This looks like it's just the normal LSST bandpass we normally use. Am I missing something here?

This retrieves the "standard passband" from the butler, which is a calibration product. See the butler code here. This was mentioned in a slack thread. The one we've been using for the visit level sims can also have per-detector throughputs included, which we wouldn't use for the coadds.

rmjarvis · 2026-10-15T14:46:45Z

+        """
+        if 'noise' not in base['image']:
+            base['image']['noise'] = { "type": "Gaussian",
+                                       "sigma": self.noise_level }


Why this bit? I think it would be safer to just make the user include this in the config if they want gaussian noise.

The object rendering needed to have a noise level in order to have reasonable stamp sizes. I actually wasn't sure if this was the right thing to do, but it seemed to make sense to have a noise level that was informed by the original image. I'm open to alternative ways of handling this.

I think it would be ok to raise an exception if the noise is missing. Could even have a custom noise type that is just CoaddGaussianNoise or something that would do this, which we could recommend in the error message.

But I think automatically adding this if it's missing is too much magic sauce. Better to be explicit IMO.

rmjarvis · 2026-10-15T14:49:15Z

+            base['image']['noise'] = { "type": "Gaussian",
+                                       "sigma": self.noise_level }
+        if self.add_noise:
+            AddNoise(base, image, current_var, logger)


Also, I think if we're doing this here, then we don't need the self.add_noise attribute. That is only used in the regular LSST_ImageBuilderBase class because FullFocalPlane doesn't add the noise for the second pass of a 2-pass image generation.

rmjarvis · 2026-10-15T14:52:26Z

+    image_pos = base['image_pos']
+    celestial_coord = base['wcs'].toWorld(image_pos)
+    ra = celestial_coord.ra / galsim.degrees
+    dec = celestial_coord.dec / galsim.degrees


I'd recommend leaving these as the celestial_coord and make that the argument to getPSF. The whole point is to avoid implicit definitions of angles being in degrees at various places in the code. Only when finally something requires degrees or radians, convert there.

rmjarvis · 2026-10-15T14:55:30Z

+            xsize = ysize = 32
+        else:
+            # Determine the stamp size from the object flux.
+            noise_var = deep_coadds.getNoiseVar(ra, dec, band) / 3.0


This / 3.0 probably deserves an explanation.

This is something else I was a bit unsure about. If I just use the noise from the image, some stamp boundaries are apparent in some galaxy renderings, especially ones that are long and thin. This factor mitigated the problem for the test image I've been using, but I don't know that it would work in general. I'm open to suggestions.

The variable itself (named noise_var) should probably not take the /3, since that feels like lying to ourselves. :) But we could document when we use it that we make stamps large enough that the profile at the edge drops to 1/3 the noise level to avoid apparent edges.

rmjarvis · 2026-10-15T15:00:31Z

+        # Probably a little better to put them a bit later than the
+        # start in some cases (e.g. after TimeSampler,
+        # PupilAnnulusSampler), but leave that as a todo for now.
+        photon_ops = psfs + photon_ops


I believe all this psf as photon_op handling is currently done correctly in GalSim itself. I think this goes back to when this was all new and we didn't have all that implemented correctly. It's probably ok, but I think it likely makes our code more fragile than it needs to be. If we can skip that and just let GalSim do it, that would be better. (Same with where we do this in the regular LSST_Image class.)

Would the fix be just to omit this line?

Also to draw the original profile, rather than split out the psfs. So skip the line where we make gal, *psfs from the original profile, and just use prof where we currently use gal. I think that should all just work, but definitely should test it to make sure I'm not missing something.

Hi Mike, I think I've addressed all of your comments. Please have a look when you get a chance.

…rdinate object instead of ra, dec

…ise type explicitly in the config

…rent stamp edges for long, thin galaxies

… as photon_ops explicitly

rmjarvis

LGTM

jchiang87 and others added 23 commits September 8, 2026 14:29

Add interface to deep_coadds and cell-based coadd PSFs

d61885b

rename coadd_psf.py -> deep_coadd.py

c4cb7ab

provide info on deep_coadds more generally

489317e

initial implementation of coadd image builder

e160fd2

initial implementation of coadd wcs builder

fffa020

add RubinDeepCoaddOutput type; adjust various configs

ef0fae0

add deep-coadd template yaml

ac0a3da

initial stamp and noise code

fc753cf

add RubinDeepCoaddBandpass

8d15b6c

implement .draw method for RubinDeepCoaddStampBuilder

d95c817

use LSST_SiliconBuilder as base class; enable fft rendering

bcd33c2

enable iterating over coadds

bedafcc

revert inadvertant changes

5917ee6

revert inadvertant whitespace change

909073b

move psf builder to a separate file; try to use consistent naming thr…

27fc58f

…oughout

switch to using lsst.images version of deep_coadd throughout; also co…

a991c70

…nsider band-dependence in the internal caches; fix pixel cell center calculation

add access to number of visits per cell; refactor caching

cc1e198

use photon ops weighting to scale object fluxes so that output units …

86c42e3

…are nJy

add default input.deep_coadds config

523ccbe

set maximum number of simulated photons to avoid excess memory usage

5fddd54

use the noise realization to estimate the noise variance for stamp si…

3beeafa

…ze calculation

use maxN in drawImage to limit the memory usage for bright objects

2edcc27

add class docstring and some comments

29b5643

jchiang87 marked this pull request as ready for review September 25, 2026 14:53

jchiang87 requested a review from rmjarvis September 25, 2026 14:54

jchiang87 added 2 commits October 12, 2026 14:14

disable fft option for deep_coadd sims

729d09a

set noise_var to avoid apparent stamp edges; add output.truth section…

02eda92

… to deep coadd config

rmjarvis requested changes Oct 15, 2026

View reviewed changes

jchiang87 added 5 commits October 16, 2026 11:04

update DeepCoadds.getPSF and .getNumVisits interfaces to pass the coo…

58610d5

…rdinate object instead of ra, dec

add RubinDeepCoaddGaussianNoise type and require user to provide a no…

dd3fb79

…ise type explicitly in the config

divide noise_var value passed to get_stamp_size by 3 to suppress appa…

729e56a

…rent stamp edges for long, thin galaxies

let GalSim handle the psfs instead of extracting them and adding them…

31daf28

… as photon_ops explicitly

tweak class docstring

3fa99a1

jchiang87 requested a review from rmjarvis October 16, 2026 22:57

rmjarvis approved these changes Oct 17, 2026

View reviewed changes

jchiang87 merged commit ab9ab14 into main Oct 17, 2026
2 of 4 checks passed

jchiang87 deleted the u/jchiang/rubin_deep_coadd branch October 17, 2026 04:55

Conversation

jchiang87 commented Sep 24, 2026

Uh oh!

rmjarvis commented Sep 25, 2026

Uh oh!

jchiang87 commented Sep 25, 2026

Uh oh!

rmjarvis left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rmjarvis left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants